ABSTRACT. Science is fundamentally transforming our lives, with technologies such as artificial intelligence (AI), gene editing, and quantum computing having profound impacts across society. However, the journey of scientific ideas from conception to societal implementation involves complex socio-technical interaction mechanisms that require systematic understanding (Roco, 2020). While AI has been studied since the 1950s, only recently has it witnessed unprecedented public interest and practical implementation, exemplified by the emergence of generative AI systems like ChatGPT that are rapidly reshaping both daily life and industry. This contrast between AI's long academic history and its recent explosive societal impact raises crucial questions about the mechanisms of scientific knowledge diffusion and social acceptance - a phenomenon that exemplifies the complex convergence-divergence patterns typical in the evolution of transformative technologies (Roco, 2020). This study investigates the dynamic interaction between scientific knowledge and social discourse in the field of AI by analyzing conceptual structure of multi-layer network. By quantitatively examining how scientific concepts interact with and evolve alongside social discourse, we aim to uncover the underlying mechanisms that drive the translation of scientific ideas into societal impact. Our study analyzes an extensive dataset spanning from 1980 to 2020, utilizing over 10 million journal articles and conference papers from OpenAlex and over 800,000 news articles from ProQuest. To understand the conceptual structure evolution in both domains, we constructed co-occurrence based semantic networks (Kedrick et al., 2024). For scientific knowledge, we utilized the hierarchical Fields of Study (FoS) relationships in OpenAlex to build two distinct networks: a convergence network that connects Level 1 FoS through shared Level 2 FoS, and a divergence network that links Level 1 FoS through individual papers. The temporal analysis of these networks reveals distinct evolutionary patterns. The social subject network shows three major peaks of activity around 1998, 2008, and 2017, characterized by balanced growth in both nodes and edges, suggesting organic expansion of social discourse. These peaks appear to coincide with significant AI milestones that sparked widespread public interest. The 1998 peak aligns with IBM's Deep Blue defeating chess champion Garry Kasparov (1997). The 2008 peak follows the introduction of the term "Deep Learning" by Geoffrey Hinton's team in 2007, marking AI's expansion beyond purely technical domains into various social sectors including transportation, finance, and industry. The 2017 peak reflects the societal impact of AlphaGo's victory over Lee Sedol (2016), which dramatically demonstrated AI's potential to the general public and triggered increased investment in AI across both private and public sectors. The scientific concept networks display notably different dynamics: while the number of nodes tends to converge over time, the number of edges continues to increase, indicating that AI's scientific evolution is characterized more by deepening connections between established concepts than by the emergence of entirely new ones. Particularly interesting is the divergence network's strong growth in new edges after 2015, suggesting an acceleration in cross-disciplinary applications of AI concepts. This surge coincides with the widespread adoption of deep learning technologies across various domains, catalyzed by landmark achievements like AlphaGo and, more recently, the emergence of generative AI systems like ChatGPT that have begun reshaping daily life and industry practices. To evaluate the structural evolution of these networks, we employ three key metrics: core churn rate (C), relative number of core nodes (R), and number of cores (S) (Kedrick et al., 2024). The social subject network shows increasing core churn and decreasing relative core size over time, indicating dynamic restructuring of central concepts in media discourse. Conversely, both convergence and divergence networks in the scientific domain exhibit decreasing core churn and increasing relative core size after the mid-1980s, suggesting the consolidation of fundamental AI concepts. Notably, while the number of cores increases across all networks, their concentration remains relatively stable around 0.1, indicating the absence of dominant core concepts. This suggests that both scientific knowledge and social discourse in AI maintain a distributed, multi-focal structure rather than converging around a single dominant paradigm. These findings provide novel insights into how scientific knowledge and social discourse co-evolve, particularly in rapidly evolving fields like AI. The contrasting patterns between scientific and social networks—particularly in conceptual structure dynamics—suggest different mechanisms of knowledge evolution in these domains. While scientific knowledge shows a pattern of conceptual consolidation with increasing interconnections, social discourse exhibits more dynamic restructuring of core concepts. Understanding these patterns and mechanisms could inform strategies for more effective science communication and technology transfer, while also providing a methodological framework for studying knowledge evolution in other scientific domains.

Kedrick, K., Levitskaya, E., & Funk, R. J. (2024). Conceptual structure and the growth of scientific knowledge. Nature Human Behaviour, 8, 1915-1923. https://doi.org/10.1038/s41562-024-01957-x Roco, M. C. (2020). Principles of convergence in nature and society and their application: from nanoscale, digits, and logic steps to global progress. Journal of Nanoparticle Research, 22, 321. https://doi.org/10.1007/s11051-020-05032-0

ABSTRACT. 1. Introduction. Emerging technologies, characterized by radical novelty, rapid growth, coherence, significant impact, and inherent uncertainty (Rotolo et al., 2015; Small et al., 2014), challenge traditional STI policy frameworks. Among these, generative artificial intelligence (AI) has become a prime example, driven by its transformative potential and rapid development. The introduction of the Transformer model by Vaswani et al. (2017) marked an important milestone, driving generative AI towards widespread adoption and interdisciplinary applications. Since then, this technology has demonstrated profound societal impacts and accelerated research worldwide, especially in the United States and China. However, other countries, such as Japan, have not experienced the same pace of growth, raising concerns about their absorptive capacity. Absorptive capacity, as defined by Cohen and Levinthal (1990), refers to the ability of a research system to identify, assimilate, and apply new knowledge. In the context of emerging technologies, "national research absorptive capacity" encompasses the ability of research ecosystems to integrate innovative research findings and transform them into impactful advances. This study extends the concept to examine how resource allocation, interdisciplinary collaboration, and institutional structures influence a nation's ability to advance generative AI research, using Japan as a case study.

2. Methodology To achieve its objectives, this study combines bibliometric analysis and survey research. The bibliometric analysis mapped the landscape of generative AI research in Japan, addressing the challenge of identifying relevant works across different fields. The seminal paper "Attention is All You Need" (Vaswani et al., 2017) served as the core reference, and citation chaining was used to define three data sets. The core set included 18,746 papers that directly cited the core reference, the marginal set included 66,802 papers that cited papers in the core set, and the boundary set included 126,120 papers that cited papers in the marginal set. Together, these datasets represented 211,668 papers, including key works such as BERT, Vision Transformer, and GPT-3. The survey targeted 575 Japanese researchers identified from the bibliometric dataset, yielding 63 valid responses (16.9% response rate). Respondents were asked about their research timeline, including when they started generative AI work, published results, and recognized its societal potential. They also rated whether their research activities expanded rapidly or gradually, and assessed their ability to adapt to the global acceleration of generative AI research.

3. Results and Findings The bibliometric analysis revealed several trends in Japan's generative AI research ecosystem. In the early stages (2017-2018), Japan demonstrated moderate absorptive capacity, ranking seventh in the world in terms of publication output. However, this was followed by stagnation in publication volume and impact metrics. Leading institutions included the University of Tokyo (600 papers), Kyoto University (239), and Tokyo Institute of Technology (233). Early adopters, such as the National Institute of Information and Communications Technology, contributed disproportionately in the early stages, with 11 papers in the first year. This disparity suggests a mismatch between early research activities and subsequent institutional growth, as larger universities entered the field later but failed to capitalize on the momentum created by early contributors. The survey results provide further insight into the dynamics of generative AI research in Japan. Approximately 20 percent of respondents reported starting their research before June 2017, when the Transformer model was introduced. After this milestone, new researchers entered the field at a steady, modest pace. By the end of 2020, approximately 30% of respondents recognized the transformative potential of generative AI, with the largest spike in awareness occurring between January 2022 and May 2023. In contrast to the global trend, where the number of new researchers grew rapidly, Japan's participation expanded gradually, driven primarily by existing researchers rather than new entrants. Regarding the scalability of research activities, 17% of respondents reported rapid expansion, 53% described gradual growth, and 30% noted no significant changes. Respondents assessed their adaptation to the global acceleration of generative AI research, with 4% believing they have adapted very well, 40% believing they have adapted adequately, and 47% responding neutrally. These results are consistent with the gradual growth trends observed. While some researchers adapted successfully, a significant proportion did not see significant change or remained uncertain about their ability to adapt. Key enablers for successful adaptation included prior computational and AI expertise and access to domain-specific datasets, which significantly influenced scalability. However, significant barriers included limited computational resources, talent retention challenges, and inadequate administrative and infrastructure support. Many skilled researchers migrated to the private sector, attracted by better compensation and stability, reducing the capacity of academia to sustain generative AI research. A ridge regression analysis was conducted to explore the factors that enabled adaptation. Missing data were imputed using multiple imputation, resulting in five data sets for analysis. The final model achieved R2 values ranging from 0.400 to 0.502. Key predictors included early engagement in generative AI research, access to key datasets, specialized expertise in AI, secured research support systems, and interdisciplinary research teams. These findings underscore the importance of foundational resources, expertise, and institutional support in enabling early-engaged researchers to effectively scale their activities and adapt to global trends.

5. Discussion This study provides insights into the trajectory of generative AI research in Japan, highlighting factors that enabled or constrained researchers' adaptability to global trends. Researchers who engaged in generative AI activities early on, supported by adequate expertise, access to essential data, robust research support systems, and interdisciplinary collaboration, were better positioned to rapidly expand their activities. However, systemic challenges such as limited computational resources, talent shortages, and insufficient institutional support remain significant barriers. Generative AI highlights broader challenges in fostering emerging technologies within national STI systems. The findings underscore the importance of maintaining a core group of researchers with absorptive capacity in nascent fields and dynamically supporting them when growth signals emerge. Potential strategies include consolidating talent in dedicated research centers, providing rapid access to computational resources, and shifting from individual to team-based funding mechanisms. By addressing these challenges, Japan can improve its ability to adapt to rapid technological advances and ensure its competitiveness in the global research landscape.

ABSTRACT. In this paper, we identify the emerging digital automation technologies and areas of scientific advance (or scientific fields) since 2010 from patented innovations and scientific publications. We cover a large spectrum of digital automation technologies, beyond robots and AI, including for example data acquisition and data management technologies, computing, networking, additive manufacturing, and user interfaces (Savona et al., 2022).

We extract a corpus of digital automation patents since 2001using a combination of Derwent codes, IPCs and keywords in the Derwent Index. Using those patents’ titles and abstracts we build queries to search the OpenAlex publication repository for scientific publications on digital automating technologies and extract a corpus of publications from 2001. Next, from each of the two corpora we identify novel and established patents and publications using an anomaly detection algorithm (Jeon et al., 2022). Novel, are the documents in the top 90th percentile of the novelty measure. These are patents/publications that create substantially different inventions, and which are most likely to introduce new technological directions or new applications within existing technologies. Established, are the documents in the bottom 10th percentile of the novelty measure. These are patents/publications that build on previous inventions and create more inventions of the same type. We also identify the documents that are most similar to those novel and established patents and publications, i.e. which develop the technology in similar directions in the future years — their ‘offshoots’. This allows to exclude patents that do not generate follow up (similar) inventions.

Next, for each set of established and novel documents we cluster them in 500 clusters of similar inventions and research areas. We obtain a large number of granular technologies, ranging from neural networks and self-driving vehicles, to block- chain and medical monitoring devices. We classify each granular technology by their pattern of growth to distinguish the speed at which they grow, the year in which they start growing, when they become established and when most of the novel patents emerge. We finally cluster all 1000 granular technologies in a smaller number of 80 digital automation technologies.

Each of these 80 technologies includes granular technologies that can be either novel or established, and which may have been following different patterns of growth. To identify these patterns, we compute the average year of the bottom 10% of the patent applications by year. We next study the temporal trends between the emergence of the different technologies. We estimate whether novel groups generally precede or follow established groups even within a broad technology. We find that novel groups, on average precede the established technologies. This suggests that a set of radically novel inventions can lead to the creation of a new technology followed by a series of innovations and improvements, resulting in the consolidation of that technology.

We distinguish three types of technologies. • First type are technologies that were novel in 2011-2015, grew exponentially in the number of inventions, forming established technologies by 2021. AI, Blockchain, IoT, Cloud computing and AM all pertain to the first group. • In the second type are technologies that were established in 2011-15 (build on technologies already common since 2000), and which gave way to novel applications (rather than technologies) through 2016-2021. Old and new ICTs related to payment, workflow management, information processing are the main technologies in this group. • In the third type are technologies that fluctuate between established and novel granular technologies more than once between 2011-2021. These are mainly robots. These technologies are relatively similar to technologies developed since 2000, but also create several novel technologies since 2011.

We finally examine the trajectories for novel and established technologies. We find that established technologies typically build upon the novel ones before 2012. These older technologies likely follow a path-dependent pattern, where initial developments create a foundation upon which newer, complementary innovations can emerge. These include more mature technologies of type 2 such as mobile and point-of-sale payment systems, messaging, digital advertising, gaming, logistics, and management systems. In contrast, technologies post-2015 (type 1) seem less path-dependent, beginning as radical innovations before moving toward incremental refinements. These include technologies cantered around autonomous and electrical vehicles, renewable energy forecasting, augmented reality, facial recognition, deep learning, cloud computing, blockchain, internet of things, additive manufacturing, and robotics. For this set of technologies, radical novelty potentially leads to establishment of the technology after a few years. In addition, some of the novel technologies such as quantum computing or artificial intelligence for health management might not have yet stabilized into an established group of technologies. Therefore, these exist only among the novel groups in our data.

Our paper makes several contributions. First, we combines the literature on emerging technologies (Rotolo et al., 2015), and on digital automation technologies (Van Roy et al., 2020; Baruffaldi et al., 2020; Martinelli et al., 2021; Singh et al., 2021; Mateos-Garcia & Klinger, 2023). We contribute by integrating and operationalising attributes of emerging technologies to a corpus of documents that identify a broader set of technologies that included in previous literature. This allows us to not only to identify what are the main technologies and their main trajectories, but also to analyse which ones are most likely to generate more radical innovations in the near future. We also contribute to the literature that uses NLP techniques for mapping and forecasting technologies (Hain et al., 2022; Arts et al., 2023; Shibayama et al. 2021; Yin et al. 2023; Kelly et al. 2021). Our approach is closest to Jeon et al. (2022). With respect to them, we use a pre-trained sentence transformer that has shown the best performance in understanding the context of written text, so far: MPNet (Song et al., 2020). Finally, we contribute to the literature that identifies technologies related to occupations (e.g. Webb, 2020; Kogan et al., 2021). We contribute to this literature by identifying a comprehensive set of 500 narrow and 80 broad digital automation technologies, applications and research areas which can be used to identify the relation between specific digital automation technologies and work.

ABSTRACT. This submission can be related to Thematic Panel: "AI in Science, Technology, and Innovation Policy Studies – Paradigm Shift or Necessary Compromise?”

----- Abstract

This work investigates the use of Large Language Models (LLMs) to identify innovations from web-scraped content, focusing on AI adaptation in Finland. The primary aim is to explore how advanced AI methods can support innovation measurement through unstructured data analysis. To achieve this, the study uses GPT-4o, a long context LLM, to extract relevant artifacts from web content, with a focus on entity identification and relationship extraction to generate knowledge graph (KG) structures. This research aims to understand how the combination of LLMs and KGs can provide a more comprehensive view of innovation landscapes.

Preliminary findings indicate that LLMs effectively capture complex innovation-related information that traditional methods may overlook. However, LLM bias toward over-identifying artifacts poses challenges, which are addressed through additional filtration steps using LLM-as-a-judge evaluations and expert review. The results underscore the potential of LLMs to enhance innovation detection and measurement at scale, while also highlighting the need for human oversight in the process. This study contributes to the growing field of LLM integration in research processes, offering insights into how such data may be evaluated and adopted for use by innovation policymakers and strategic managers.

Introduction

Innovation is critical for economic growth and competitiveness. However, traditional methods for measuring innovation, such as surveys and patent analyses, often lack comprehensiveness and fail to capture emerging trends promptly. These traditional indicators frequently lack sectoral, technological, and geographical detail, leading to a narrow, retrospective view of innovation activities. The rapid advancement of digital technologies and the complexity of innovation ecosystems create a need for more adaptable and scalable approaches to measure innovation efficiently.

This work addresses these limitations by leveraging LLMs to extract innovation indicators from web-scraped content, providing a new approach for identifying and classifying innovations. LLMs, such as GPT-4, can perform high-level analysis of unstructured text-based data, making them well-suited for sources like company websites, news articles, and industry reports. Our research aims to answer the following questions: (1) How can LLMs effectively extract innovation indicators from web-scraped data? (2) What advantages do LLMs offer over traditional methods? (3) How can integrating LLMs with KGs enhance the analysis of unstructured data?

Methodology

The study uses web-scraped data from company websites, industry reports, and other publicly available sources, processed using LLMs to extract relevant entities and relationships. A variety of web pages were scraped, including corporate websites, press releases, and research publications, to ensure a diverse dataset. Data preparation includes cleaning, deduplication, and structuring the extracted information into KGs. LLMs facilitate the extraction of key information, such as types of innovation, thematic focus, and organizational involvement, even from unstructured and noisy sources.

Once the entities and relationships are extracted, the information is structured into KGs. KGs are used to visualize and analyze the relationships among entities, such as companies, products, technologies, and collaborations, providing insights into the innovation ecosystem. By linking these entities, KGs offer a deeper understanding of how different components of the innovation system interact and evolve.

The methodology also includes validation by comparing extracted information with existing innovation databases to assess the accuracy and reliability of the LLM-based approach, ensuring that the insights complement traditional innovation metrics.

Preliminary Findings

Initial results indicate that LLMs are promising in detecting innovation-related information from large datasets, revealing insights that traditional methods may miss. However, even with detailed instructions, LLMs tend to over-identify instances of innovation, leading to inflated results. To mitigate this, we incorporated additional filtering mechanisms, including a secondary evaluation step using LLM-as-a-judge techniques and expert reviews. These steps help reduce biases and improve reliability, but highlight that fully automating the process still requires human oversight and intervention.

We also found that the integration of KGs enhances the representation of complex relationships, providing deeper insights into the innovation landscape. KGs reveal connections between innovations, such as linkages between a company's R&D activities and partnerships, offering a richer understanding of the broader innovation ecosystem.

Another key finding is the ability of LLMs to identify patterns and trends in innovation across sectors. For example, the analysis revealed a growing focus on sustainability-related innovations, with many companies investing in green technologies. This insight is crucial for policymakers and industry leaders to understand the direction of technological development and allocate resources accordingly.

Discussion and Conclusion

The findings suggest that LLMs, when combined with KGs, hold significant potential to improve the measurement and analysis of innovation. LLMs address several limitations of traditional methods, such as their inability to capture real-time developments and reliance on structured data. By analyzing unstructured web data, this study demonstrates how AI integration can offer a more dynamic and up-to-date view of innovation activities. Moreover, KGs provide a structured representation of the innovation ecosystem, helping to identify key players, emerging technologies, and potential areas for collaboration, which is particularly valuable for informed decision-making by policymakers and industry leaders.

Future research should focus on refining LLM models to reduce issues such as over-identification and hallucinations and expand the dataset to include more diverse information sources. Integrating other AI techniques, such as machine learning-based clustering and classification, could further enhance innovation data analysis. Exploring hybrid models that combine LLMs with other AI approaches could also lead to more robust and reliable insights.

ABSTRACT. Abstract

This work explores the integration of Large Language Models (LLMs) into an international innovation policy database (EC-OECD STIP Compass), a system that aggregates data on global science, technology, and innovation (STI) policies. As the volume of data grows, traditional methods for collecting and validating survey responses have proven labor-intensive. To address this, we conducted an experimental pilot using GPT-4o, a long-context LLM, to evaluate whether LLMs can assist in prefilling survey data by extracting relevant information from web-scraped content. The findings suggest that LLM-assisted data collection can significantly enhance efficiency, but challenges remain regarding accuracy and information overlap. This research lays the groundwork for future refinements in the use of LLMs for wide policy data collection and validation.

Introduction

Limited availability of comprehensive and comparable data across countries and regions is a major challenge in understanding the design and effects of innovation policies. This limitation arises due to the costs of collecting such data and challenges related to the unit of analysis—as policies do not easily lend themselves to large-scale data collection (Flanagan et al. 2011). Large Language Models are suited to process complex, unstructured data and are highly scalable, potentially addressing these challenges (Feldman et al. 2015). However, LLMs are not a silver bullet for innovation policy studies. They require input data reflective of the real-world distribution of innovation policies, which cannot be inferred from publicly available data without human expertise. Additionally, LLM-driven data-generation processes introduce noise into the data (Shumailov et al. 2024).

In this paper, we propose a methodology that leverages the strengths of LLMs for data-generating processes while safeguarding against their weaknesses. Our method integrates LLM usage with an expert survey, reducing data collection costs by minimizing experts' time while ensuring accuracy in the LLM-generated data.

The use of LLMs, which have shown promise in other domains of information extraction and text analysis, offers a potential solution to streamline data collection processes. LLMs can process large amounts of text quickly, enabling more efficient data extraction from web-based sources, which could augment traditional human-led surveys. This study focuses on a pilot experiment designed to assess the effectiveness of LLMs in assisting with the EC-OECD STIP Compass survey.

LLMs, particularly models such as GPT-4o, have been lauded for their ability to handle long-form text and perform complex natural language processing (NLP) tasks, including information extraction, summarization, and question answering. However, their application in policy data collection—particularly in STI policy—remains relatively underexplored. The primary goal of this study is to evaluate how effectively an LLM can fill in survey responses based on content scraped from URLs provided by human respondents.

Methodology

The experiment was conducted using a pilot sample of six countries from the OECD—Canada, Finland, Germany, Korea, Spain, and Turkey. The methodology focused on testing two primary approaches to LLM integration: (1) Retrieval Augmented Generation (RAG) and (2) LLMs with Long Context Windows, specifically GPT-4o.

The experiment involved the following steps:

Data Preparation: Survey data was drawn from responses provided by human participants, which included relevant URLs for each policy initiative. Web scraping was performed to collect the full content of the URLs. Initiatives with fewer than 200 tokens or more than 120,000 tokens were excluded from analysis.

LLM Integration: Key survey questions were used to design prompts for GPT-4o. For web content exceeding the token limit, chunking methods were applied. The LLM was prompted to summarize and extract relevant information on policy instruments, objectives, target groups, and policy themes.

Comparison and Evaluation: The LLM outputs were compared to human responses. We evaluated the overlap between human-provided and LLM-generated responses across dimensions such as policy instruments and objectives. The methodological process includes a validation layer that assesses LLM performance (e.g., flagging hallucinations).

Findings

The experiment produced mixed results, highlighting both the potential and limitations of LLMs in this context. In terms of overlap, the LLM captured at least one relevant policy code (e.g., target group, instrument, or theme) in 95% of cases. This result indicates that the LLM was able to identify key themes and instruments that aligned with the policy content.

However, at more detailed levels of comparison, such as full descriptions and objectives of the policy initiatives, the overlap was much lower. Only 1.19% of descriptions had full overlap between LLM and human responses, with 74.05% showing high overlap and 9.52% showing no overlap at all.

A deeper analysis revealed that low overlap was often due to differences in how the LLM and humans prioritized information. Human responses emphasized practical implications and societal impacts, while the LLM focused more on technical details or procedural aspects.

Discussion and Conclusion

This pilot study demonstrates that LLMs can be valuable for prefilling policy survey responses, offering significant advantages in terms of time efficiency and flexibility. However, important limitations must be addressed before LLMs can be fully integrated into large-scale data collection efforts like the EC-OECD STIP Compass.

One key challenge is developing effective validation mechanisms to catch hallucinations and incorrect responses. This could involve using multiple LLMs in tandem, where one model evaluates the output of another, or employing human oversight where LLM-generated responses are unclear or potentially inaccurate. Improving the quality and relevance of input data is also crucial. As shown, LLM performance depends on the quality of the web-scraped content. Ensuring only high-quality URLs are used, or incorporating additional preprocessing steps to filter out irrelevant content, would likely improve LLM performance.

The application of LLMs in STI policy data collection presents a promising avenue for improving survey efficiency and effectiveness. While the pilot results are encouraging, significant work remains to address LLM limitations and optimize their performance. Future research should focus on refining prompt design, enhancing validation methods, and improving input data quality. With these improvements, LLMs could play a transformative role in policy data collection, validation, and analysis, providing policymakers with more timely and comprehensive insights into the global innovation landscape.

ABSTRACT. Thematic Panel: AI in Science, Technology, and Innovation Policy Studies – Paradigm Shift or Necessary Compromise?

Panel Overview Artificial Intelligence (AI) has rapidly transformed the landscape of science, technology, and innovation (STI) studies. Its profound capabilities are reconfiguring the way we collect, analyze, and interpret data, leading to breakthroughs in STI policy and research methodologies. As we gather at the 10th Biennial Atlanta Conference on Science and Innovation Policy, our thematic panel will dive into how AI is reshaping the contours of STI studies, explore the opportunities it presents, and examine the challenges that come with its integration into research frameworks. This discussion is anchored by guiding questions on AI literacy, transparency, and its effects on both the quality and trajectory of STI research.

The panel seeks to address pressing questions about AI’s role in the study of innovation and technology policy. Our guiding inquiries are: How does AI change the nature of STI studies? Is the infamous “AI black box” a necessary trade-off for innovation, or is it merely the beginning of a new, transparent paradigm in research? Furthermore, what kind of AI literacy is essential to ensure that STI studies remain scientifically sound, ethically grounded, and impactful? The integration of AI into STI research promises an era of immense potential, yet it also raises important issues of trust, accountability, and epistemic reliability. These questions are fundamental as we explore the transformative nature of AI-driven methods in policy research—methods that are increasingly being adopted globally in industry, academia, and government alike.

Presentation Summaries Firms’ Knowledge Disclosure: Website, Publication, and Patent Data This presentation will delve into the changing landscape of corporate knowledge disclosure through various mediums, including websites, academic publications, and patents. By leveraging AI and large datasets, the presenter will illustrate how firms use these avenues to communicate their innovation activities and how AI can enhance the extraction and comparison of disclosed information. The talk will focus on how AI contributes to building a comprehensive picture of innovation by examining patterns of disclosure that are otherwise complex and dispersed. AI-Driven Innovation Measurement: Testing the Limits of Large Language Models (and Knowledge Graphs) for Scaling the Mapping of Business Innovations Large Language Models (LLMs) are at the heart of the evolving AI toolkit for STI analysis, but how far can they really go in automating innovation measurement? This presentation will explore the integration of LLMs with knowledge graphs to enhance our understanding of business innovation activities. It will critically assess the ability of LLMs to map firm-level innovations at scale, presenting both the opportunities and challenges encountered in testing these technologies for real-world STI policy applications. What are the limits, and how do they inform the future trajectory of innovation measurement? Large Language Models in Data Generating Processes for Innovation Policy Studies: Method Development for an International Database This presentation will discuss the development of methodologies for generating data that informs innovation policy, focusing on international databases. By using AI, particularly LLMs, researchers can create consistent, cross-country datasets that provide richer insights into science and technology trends. However, challenges around transparency, bias, and reliability of data generation will be examined to highlight the complexities of using AI in building foundational policy datasets. 4. Analyzing Technologies Instead of Counting Patents! Using AI to Cluster Patents by Technologies of Enterprises Traditional methods of assessing technological innovation often involve counting patents, which may miss the nuances of actual technological advancements. This presentation proposes an AI-based alternative: clustering patents according to the underlying technologies they represent. By using AI to understand patent content, researchers can generate more meaningful insights into the technological strategies of enterprises, allowing policy-makers to better assess innovation dynamics and make informed strategic decisions. The presentation will reflect on the potential benefits, but also pitfalls of AI-based approaches.

Panel Discussion: The Path Ahead for AI in STI The panel will explore AI's transformative impact on STI research, focusing on challenges such as the opacity of advanced AI models and the balance between leveraging their power and maintaining transparency for evidence-based policymaking. Audience participation is encouraged to enrich the discussion on the "AI black box" and its implications for research credibility.

As AI capabilities expand, the need for AI literacy among researchers and policymakers becomes vital to ensure scientific integrity and maintain public trust in policy outcomes. The panel will discuss essential AI knowledge required for effective STI research and innovation policy.

We will also examine whether AI is ushering in a new paradigm for STI studies, pushing beyond traditional methodologies towards novel approaches for understanding innovation systems. Presentations will demonstrate how AI models, particularly LLMs, are being used for STI analysis, highlighting both their transformative potential and the complexities involved.

As we celebrate the 10th Biennial Atlanta Conference on Science and Innovation Policy, this thematic panel aims to reflect on AI's role in navigating the complexities of technology and innovation. By addressing AI literacy, transparency, and innovation, we hope to foster conversations that will shape the future of STI policy and research. Participants are invited to contribute, rethink, and advance the potential of AI for informed and effective innovation policies.

ABSTRACT. Traditional methods of assessing technological innovation often involve counting patents, which may miss the nuances of actual technological advancements. Usually, a bundle of patents protects one technology or the technological portfolio of a company. One the one hand, technology development is an evolutionary process within a company that might have a radical or at least “new-to-the-firm” innovation at one point and further improvements or technological advances of these at later stages. On the other hand, the IP protection should split the risk of a technology to be either attacked, circumvented or re-engineered by filing more than one patent document per unit of invention – be it the initial radical or just an incremental step. Technologies might have different propensities to patent, but also different propensities to split. So far, innovation statistics and innovation analytics has mostly used a patent-counting approach (one patent = one technology/invention). Conceptually, however, this can be challenged, given the reality described above. To tackle this problem, the idea is to identify bundles of patents per company (patent clusters), not only to improve the patent/innovation statistics, but to be able to compare technology portfolios in particular technological domains across companies/enterprises, in addition. Hence, this presentation proposes an AI-based alternative to simply counting patents by clustering patents per company/enterprise according to the underlying technologies they represent. As we do not have annotated documents at large scale that allow us to train a model (machine-learning), we developed an alternative approach using the LLaMA-Model to identify technologies in our corpus of worldwide patent documents (titles and abstracts). We follow a two-step approach to accomplish our task of clustering patents and to compare them across companies/enterprises. In a first step, we identify technologies in the patent corpus, which are represented as clusters of synonym words. We use PATSTAT as a source for our patent data and restrict the data to patent applications filed since 2010 at the transnational level. The whole step 1 could be seen as a topic model; the task is to identify technologies (topics) in the texts. But it's not a "standard" topic model, because we do not aim to assign a patent (text) to a technology (topic). We therefore prefer to call it "keyword extraction" and "synonym grouping" to avoid misunderstanding. This serves as a preparation for step 2. We allow that some outlier technology candidates stay unclustered. If we force to assign all technology candidates to (at least) one cluster even when the probability is low, the clustering quality gets much worse. Anyway, it is unnecessary, because our purpose in step 1 is to detect the existing technologies. It does not matter even if one header word cannot cover all possible wordings! In a second step we perform the patent clustering based on a similarity search, combined with a keyword search. For each patent of a company, we assign the binary (yes/no) information, if the patent belongs to a specific technology cluster that we identified in step 1. We use additional information like IPC class or citation links to further narrow down the potential similarity of patent documents per company. It needs to be stressed that we should not conclude "this patent belongs to the technology" from step 1. The experiment is not designed for that purpose. All we can say from this step is "in the texts, technologies xxx, yyy, and zzz are recognized". This does not guarantee the completeness (there might exist other unidentified/missed technologies), and exclusiveness between identified technologies. Our experiments do not support the causation link ("therefore"-reasoning), for example from the keyword "wind turbine" to the technology "wind energy"! In consequence, validation and plausibility checks are needed and are even crucial for approaches like this. Standard methods or procedures to conduct such checks have not yet been established. While our methodology follows a structured approach, based on conceptual thoughts (theorizing/hypothesizing) and produces reasonable results, we still cannot conclude that the results are valid. LLMs have been shown to “hallucinate” by making up things. In this case, the challenge is an assessment scheme for the validity of the results and the scrutiny of the established scientific perspectives in STI policy analyses. By using AI to understand patent content, researchers can generate more meaningful insights into the technological strategies of enterprises, allowing policy-makers to better assess innovation dynamics and make informed strategic decisions. But how reliable and valid are these insights? However, established scientific approaches might be challenged by the new methods and the new possibilities based on LLMs. As the LLMs are “black boxes” themselves, new scientific perspectives might become necessary. The presentation will use the suggested method to reflect on the potential benefits, but also on pitfalls of AI-based approaches in this context, thereby trying to refer to the three overarching questions of the thematic panel "AI in Science, Technology, and Innovation Policy Studies – Paradigm Shift or Necessary Compromise?".

ABSTRACT. Failure is an inherent and essential component of scientific progress, shaping the trajectories of individual careers and advancing the boundaries of knowledge. Scientists face setbacks regularly, ranging from minor experimental issues to major challenges such as unmet project objectives, rejected publications, or unsuccessful grant applications. These experiences often lead to undesirable outcomes, including cumulative disadvantages that can hinder career advancement over time (Bol et al., 2018; Merton, 1968). Despite its prevalence, failure is frequently associated with negative connotations, often perceived as a result of inadequate competence or insufficient effort. This stigma not only discourages open discussions about failure but also overlooks its potential to drive innovation and learning in science. Contrary to the traditionally negative perception of failure, recent empirical research highlights its potential as a powerful driver of scientific success. Studies show that setbacks can lead to significant career achievements when they are met with resilience and adaptation. For instance, research by Wang et al. (2019) revealed that narrowly missing out on prestigious grants in early careers often pushes scientists to adopt more innovative approaches, ultimately leading to greater long-term success. Similarly, Yin et al. (2019) demonstrated how early failures in funding applications could foster persistence and creativity, emphasizing the critical role of learning from failure in scientific careers. These findings challenge the traditional narrative of failure as purely detrimental, suggesting that it can be reframed as important steps to greater achievements. Recognizing the dual nature of failure, as both a challenge and an opportunity, this study aims to deepen our understanding of its role in scientific process. We focus on how risk-taking behavior influences setbacks, particularly those related to funding and publications, and how the consequences of such setbacks vary across cultural contexts. By analyzing responses to setbacks, we aim to uncover patterns in resilience and adaptability, shedding light on the mechanisms that enable some researchers to progress despite failure. The study focuses on the interplay of individual, organizational, and cultural factors, recognizing that scientists do not work in isolation but are embedded within specific environments that shape their experiences and responses to failure. Our research is framed within the context of three countries: Sweden, Japan, and the United States. These countries were chosen for their contrasting cultural attitudes toward risk and failure, as well as their substantial investments in research and development. Previous research have highlighted notable differences in how these cultures approach risk-taking. The United States is characterized by a high tolerance for risk and a culture that often celebrates failure as part of the entrepreneurial and scientific journey. In contrast, Japan is known for its more cautious approach to risk, valuing precision and incremental progress, while Sweden occupies an intermediate position, balancing risk-taking with a strong emphasis on consensus and stability (Hofstede, 2001). These cultural differences likely influence how scientists approach high-risk research and respond to failures, yet little empirical work has investigated this dynamic in depth. More specifically, our study examines the following research questions: 1. How do risk-taking behaviors in science impact the likelihood and perception of setbacks? 2. How do cultural attitudes toward risk and failure shape researchers' responses to setbacks? 3. What factors—such as career history, research type, or novelty—moderate the relationship between risk-taking and setbacks?

The data for this study is being collected through an online survey in the above countries, targeting assistant professors, associate professors, full professors, and adjunct professors. The survey explores a range of topics, including demographic and professional backgrounds, experiences with setbacks, and strategies for overcoming challenges. It also seeks to measure individual risk-taking behavior, perceptions of setbacks, and the broader organizational environment's tolerance for setbacks. To complement this survey data, bibliometric analyses are being conducted to contextualize respondents’ research histories. These analyses include information on the novelty and interdisciplinarity of their work, shifts in research focus over time, and productivity patterns. A key aspect of this research is its focus on the organizational environment, which plays a critical role in shaping scientists’ responses to failure. Supportive organizational cultures can help mitigate the negative effects of setbacks, encouraging resilience and fostering innovation. For example, institutions that provide resources for risk-taking and actively destigmatize failure can create environments where scientists feel empowered to pursue ambitious, high-risk research. Conversely, environments that penalize failure or prioritize short-term success can stifle creativity and discourage researchers from exploring unconventional ideas. By analyzing the organizational contexts in Sweden, Japan, and the United States, this study aims to identify best practices for fostering resilience and innovation within research institutions. In addition to organizational factors, this study examines the influence of broader cultural environments on scientists’ experiences with failure. Cultural attitudes toward risk-taking and resilience can shape how setbacks are perceived and managed. For instance, the relatively high risk-tolerance in the United States may encourage scientists to embrace failure as an integral part of the research process, while the more risk-averse culture in Japan might lead to greater caution in pursuing high-stakes projects. Sweden’s balanced approach may offer unique insights into how to navigate the trade-offs between risk-taking and stability. By comparing these cultural contexts, this research seeks to illuminate how national cultures influence the scientific process and the conditions under which failure can be transformative. This research contributes to the growing understanding of failure as an integral aspect of scientific progress. By exploring how scientists in different cultural and organizational environments experience and respond to setbacks, this study aims to inform policies and practices that promote resilience, risk-taking, and innovation in the global research community. Recognizing failure not as an endpoint but as a critical step in the scientific process can help foster a culture of learning, adaptability, and creativity, ultimately driving the advancement of science.

ABSTRACT. Risk and risk-taking in science and their relation to creativity and novelty in research results have received increasing attention in recent years (Franzoni and Stephan 2023). In particular, many have argued that researchers lack incentives to engage in risky research (Franzoni et al. 2022, Alberts et al. 2014). For example, institutional conditions are argued to work against risk-taking (Heinze 2013, Lee and Walsh 2022, Stage and Utoft 2023), including performance assessment based on number of publications, employment in short-term positions, reliance on competitive research funding, and conditions for advancement. Particular focus has been placed on risk-taking in the allocation of competitive funding and peer review of research grant proposals (Franzoni and Stephan 2023, Carson et al. 2023). Both the way in which researchers perceive these review processes to function and reviewers’ and funders’ willingness to fund risky research can have important impacts on what research is funded and its novelty. As also noted by Franzoni and Stephan (2023), despite this focus, there is a lack of clarity about what risk-taking in science is and how it is perceived. There may also be multiple types of outcomes related to research, such as for careers, funding and ethics, suggesting that risk should not be seen solely in terms of scientific outcomes. Risk-taking in science can also be seen from a number of different perspectives, most notably from that of funders, reviewers and researchers. Each group has their own set of factors that can determine what risk-taking is but are closely influenced by how risk-taking is perceived by other actors. This paper seeks to examine the following questions. How is risk and risk-taking perceived by researchers? What strategies do researchers use to address uncertainty in a funding program that promotes originality? The analysis is based on grant recipients from the Villum Experiment Programme . This program, which is funded by the Villum Foundation in Denmark, supports projects exploring original research ideas that are out of the ordinary, challenge established norms within one or more fields, and could potentially change the future scientific approach to a given topic (Sinkjær 2018). The analysis draws on semi-structured interviews with 24 Villum Experiment grant recipients in 2017 or 2018. The interview guide encompassed five categories of questions: the recipient, their perceptions of the funding instrument, the applied for research plans, the research process and outcomes. Our study indicates, on the one hand, all appeared to perceive that risky applications would be positively received by the program and most though not all researchers actually designed projects that had greater risk than their research in general. Researchers’ perceptions of the VE program and of how applications are assessed appeared to influence how they wrote their VE project applications and how the projects were conducted. Importantly, on the other hand, we also find many examples of how conditions within the Danish research system influenced risk-taking even in a program such as the VE, particularly concerning the potential consequences for career paths or performance evaluation. Grantees expressed risk in a number of different ways, typically not in one single way, such as in terms of chances of success, uncertainty due to use of new methods or equipment or a new research area. Risk was also expressed in terms of uncertainty concerning how long it would take to conduct planned research experiments. Outcomes were typically referred to in terms of publication potential: possible number of publications that the research could produce or risk of not being able to publish anything and the level of journals that could be reached. While scientific outcomes were often expressed in terms of publication potential, researchers were mainly concerned with potential career-related effects of a negative scientific outcome on future job opportunities, particularly for temporarily employed postdocs, or in some cases their chances of securing subsequent funding. In particular two related strategies were employed to “hedge” project risk on behalf of postdocs hired for the projects. The first was to ensure that no participants worked only on the VE project, that VE participation was balanced with participation in another research project with more certain outcomes. The second was to divide the VE project into two parts, where the first part had a certain outcome (in some cases, a literature review) and the second part was experimental with uncertain outcomes. References Alberts, Bruce, Marc W. Kirschner, Shirley Tilghman, and Harold Varmus. 2014. “Rescuing US Biomedical Research from Its Systemic Flaws.” Proceedings of the National Academy of Sciences 111 (16): 5773–77. https://doi.org/10.1073/pnas.1404402111. Carson, R. T., Zivin, J. S. G., and Shrader, J. G. (2023). Choose your moments: Peer review and scientific risk taking (No. w31409). National Bureau of Economic Research. Franzoni, C., Stephan, P., and Veugelers, R. (2022). Funding risky research. Entrepreneurship and Innovation Policy and the Economy, 1(1), 103-133. Franzoni, C., and Stephan, P. (2023). Uncertainty and risk-taking in science: Meaning, measurement and management in peer review of research proposals. Research Policy, 52(3), 104706. Heinze, T. (2013). Creative accomplishments in science: Definition, theoretical considerations, examples from science history, and bibliometric findings. Scientometrics, 95, 927-940. Lee, Y. N., and Walsh, J. P. (2022). Rethinking science as a vocation: One hundred years of bureaucratization of academic science. Science, Technology, & Human Values, 47(5), 1057-1085. Sinkjaer, T. (2018). Fund ideas, not pedigree, to find fresh insight. Nature, 555(7697), 143-144. Stage, A. K., and Utoft, E. H. (2023). Fun and less fun funding: the experiential affordances of research grant conditions. Science and Public Policy, 50(6), 1091-1102.

ABSTRACT. Life scientists in universities are subject to an increasing number of rules and regulations stemmed from within and outside. Internally, university policies and procedures are established not only to meet external regulatory demands on transparency, openness and research ethics (Archibald & Feldman, 2008; Johnson, 2020), but also to ensure their own survival and growth. Externally, universities are subject to rules from regulatory, funding and accrediting entities at multiple levels. Research investigating the effects of regulations on scientists’ research outputs has produced mixed results. Some work argues that administrative burden produced by regulations generates increased burden for scientists, which may create barriers to resource and knowledge exchange, delay or sabotage research activities, and interfere with the traditional norms of collaborative science (Bubela et al., 2015; Eisenberg, 2001; Yeh et al., 2017). These effects are generally linked to lower science production. On the other hand, some argue that the negative impact of regulations on research process is overestimated or, at a minimum, are poorly understood and deserve greater attention (Derrick & Bryant, 2013; Mishra & Bubela, 2014). Moreover, many of the rules are designed to serve important rationales such as to conserve of biodiversity, protect of property rights, human and environmental safety and health, and facilitate open (Fusi et al., 2019; Kamau et al., 2010; Rodriguez, 2005). The administrative burden theory is deemed a viable framework to understand rule burden in the context of university research (Bozeman et al., 2020), although it is originally used to examine the effect of rule burden for citizen access to public programs in citizen-state interactions (Fox et al., 2020; Heinrich, 2016; Moynihan et al., 2015). In the context of academic research, administrative burden includes the time and efforts scientists invest in learning about various requirements and specifications, completing applications and reports as well as any psychological strain they may experience in the process. The increased job demands not only add to employees’ workload, enhance role ambiguity, but also occupational stress and phycological risks, which could impair the overall quality of academic work among academic faculty (Boyd & Wylie, 1994; Pace et al., 2021; Tight, 2010). An adjacent concept to administrative burden is organizational red tape. While both administrative burden and red tape can be burdensome, they are distinct in their functionality (Campbell et al., 2023; van Loon et al., 2016). Only rules that entail wasteful consumption of organizational resources without advancing organizational goals are considered red tape (Bozeman & Feeney, 2014; Campbell et al., 2023). At the individual level, a recent synthesis of red tape studies indicates that the effect of red tape on job involvement and employee performance is negligible (Blom et al., 2021). In fact, high-quality rules and the ways in which rules are implemented can even foster rule compliance and improve job satisfaction (Borry et al., 2018; DeHart-Davis et al., 2015). In the context of academic research, some rules and regulations are designed for important public policy reasons, such as to protect human and environmental safety, as well as intellectual property (Bozeman & Youtie, 2020; Woelert, 2023). These rules, albeit burdensome, may not necessarily raise barriers to material sharing nor hinder knowledge diffusion and research progress (Mowery & Ziedonis, 2007; Stern, 2004; Walsh et al., 2003). Accompanying increased administrative rules and enhanced formalization of institutional processes, universities worldwide are experiencing the expansion of bureaucracy (sometimes “administrative bloat”) characterized by increased administrative positions and expenditure, as well as unbundling of functions, services and faculty roles (Macfarlane, 2011; McCowan, 2017; Ramirez & Christensen, 2013). In terms of research management, professional staff are found to have a positive role in securing and managing funds through informing project designs, providing trainings and grounding rules on proposal writing for researchers (Beime et al., 2021; Ito & Watanabe, 2020). To understand whether the ways in which regulations are expressed and realized influence academic scientists’ research performance, this study examined the two aspects of rule: rule burden and rule legitimacy drawing on insights from both the administrative burden and red tape literature. We hypothesize that rule burden is negatively related to research performance and rule legitimacy is positively associated with research performance. We also hypothesize that organizational support in regulation management has a moderating role between rule legitimacy and future research performance. To test our hypotheses, we use a unique database combining survey data and bibliometric data retrieved from the Scopus database in 2024. The survey was administered in 2016 and 2017 to a national sample of 3,933 tenured and tenure-track academic scientists in the United States in three disciplines—marine biology, entomology, and ecology—from research universities classified as having “very high” and “high” research activity as defined by the 2010 Basic Carnegie Classification. We measure rule burden, rule legitimacy and organizational support using survey data.The cumulative number of publications and citations in the four years after the survey are used as dependent variables measuring research performance. Using zero-inflated negative binomial model, we found that neither perceived rule burden nor rule legitimacy is not predictive of academic scientists’ research performance. Additionally, support from university professional staff is positively associated with future research performance. This article contributes to the understanding of the relationship between rule and research performance of academic scientists by connecting the public administrative literature on rules and red tape to science policy literature. Our findings suggest that although rules and regulations may increase workload for academic scientists, they do not necessarily hinder research performance. The compliance burden can be compensated by the legitimate causes rules may serve. Moreover, challenging job demands derived from increased rule compliance burden might provide opportunities to enhance proficiency in job-related tasks, and to demonstrate competences that can get rewarded (Crawford et al., 2010). The findings also suggest that research universities should invest in specialized personnel to assist scientists in navigating the complex regulatory environment.

ABSTRACT. Motivation The global circulation of the scientific workforce is a well-established phenomenon that plays a critical role in the generation and dissemination of knowledge. It also serves as a catalyst for innovation. As globalization intensifies, scientists are becoming more mobile than ever before. The international relocation of scientists, who carry with them unique expertise and knowledge, strengthens research systems by fostering knowledge exchange across universities and regions. While debates persist about the impact of mobility on individual career advancement, there is broad consensus that mobility benefits the research ecosystem as a whole. Numerous policies at various levels actively promote the mobility of scientists to enhance the efficiency and productivity of their respective research systems. However, most research on scientist mobility has focused on what might be considered "normal" periods of scientific development. The effects of global exogenous shocks, such as a pandemic, on the international mobility of scientists remain an open and underexplored question. This study seeks to address this gap by examining how the COVID-19 pandemic has influenced the international mobility of scientists, focusing specifically on the field of Computer Science. We aim to assess whether and in what ways the pandemic disrupted traditional patterns of international mobility. Additionally, this study investigates the applicability of the traditional Push-Pull model in explaining the movement of scientists during this unprecedented global crisis. Specifically, we explore the immediate impact of COVID-19 at both the country and individual levels, seeking to understand how country-level, policy-level, and individual-level factors interacted with the adverse effects of the pandemic. Data and research method The primary dataset for this study is the DBLP-Citation-network V14 (referred to as DBLP), a comprehensive bibliometric dataset released in February 2024. This dataset, built on the original DBLP, encompasses 5.26 million papers and over 36 million citation relationships in the field of Computer Science. DBLP employs a hybrid disambiguation approach—combining algorithmic techniques with manual error correction—to achieve highly accurate author disambiguation. To extract country-level information from authors' address strings, we use a combination of string-matching algorithms and large language models. To validate the accuracy of this process, we manually reviewed 500 random address strings, achieving weighted precision, recall, and F1 scores of 0.948, 0.942, and 0.943, respectively. These results confirm a satisfactory level of accuracy in identifying country information. We use shifts in authors' affiliations across countries as a proxy for international mobility events. Bibliometric data offers an extensive, data-driven perspective on researcher mobility. By tracking changes in researchers' affiliations as reflected in their publications, we provide a comprehensive analysis of global scientific mobility. To examine the impact of the pandemic, we apply a novel causal inference approach: a difference-in-differences method with a continuous treatment variable. This method enables us to measure the disruption caused by COVID-19 on international mobility in Computer Science between 2013 and 2022. To address potential endogeneity concerns, we use the latitude of each country's capital city as an instrumental variable. Latitude serves as a valid instrument because it is correlated with temperature—which affects the transmission of COVID-19—but is exogenous to the pandemic itself. Major findings This study examines the relationship between a country's COVID-19 infection rate and the international mobility of scientists in Computer Science. Our findings indicate that an increase in a country's infection rate by one standard deviation leads to a significant reduction in population flow. Specifically, we observe a 5.35% decline in international mobility, which can be further broken down into a 3.57% decrease in the inflow of scientists and a 6.47% decrease in the outflow. These results are robust across alternative metrics, such as COVID-19 infection and death rates. To address endogeneity concerns, we employ the latitude of a country's capital city as an instrumental variable. This approach allows us to control for confounding factors that may influence both COVID-19 infection rates and scientific mobility. Our analysis reveals that the impact of COVID-19 infections is most pronounced in the second year following the outbreak. Additionally, we find that the decline in inflow is more significant in developed countries, while the reduction in outflow is more pronounced in nations with stringent anti-COVID-19 policies. These findings highlight the complex interplay between public health crises, policy responses, and global scientific mobility. Conclusion This study highlights the significant impact of the COVID-19 pandemic on the international mobility of scientists, particularly in the field of Computer Science. By leveraging a comprehensive bibliometric dataset and employing robust causal inference methods, we provide strong evidence that the pandemic disrupted established patterns of global scientific mobility, with notable declines in both the inflow and outflow of researchers. Our findings suggest that a country's infection rate is a key determinant of these mobility trends, with higher infection rates leading to reduced international movement of scientists. The differential effects observed—where developed countries experienced greater declines in inflow and nations with stringent anti-COVID-19 policies saw more substantial decreases in outflow—offer important insights into the unequal impact of the pandemic across different contexts. These results also suggest that while the traditional Push-Pull model remains broadly applicable, it must be adapted to incorporate additional factors, such as public health crises and corresponding policy responses, to fully explain mobility patterns during periods of global disruption. In conclusion, this research offers a data-driven perspective on the pandemic's disruption of the global scientific workforce and underscores the need for adaptive strategies to sustain international mobility in the face of future crises. Policymakers and research institutions should draw on these findings to develop resilient systems that support global knowledge exchange, even in times of unprecedented challenges. Future research could build upon this work by exploring the long-term recovery of scientific mobility and its implications for innovation and collaboration in a post-pandemic world.

ABSTRACT. This research examines how the EU Research and Innovation Framework Programmes (Horizon 2020 and Horizon Europe) contribute to the Digital and Industrial Transition (DIT), a pivotal element of Europe’s broader social, economic, digital, and environmental transformation. The study evaluates how these programmes address systemic challenges to foster the digital transformation of industry, leveraging a multi-level analytical framework encompassing inputs, processes, and outcomes at policy, programme, and project levels. The analysis focuses on mechanisms that enhance resilience, competitiveness, and uptake of digital technologies across sectors.

Research Questions The study is structured around three core questions: 1. To what extent do the EU R&I programme support systemic changes essential for advancing the digital transformation of industries? 2. How do key policy mechanisms, including for instance directionality and built-in flexibility in programme design and implementation, manifest in the Framework Programme to drive digital transformation? 3. What evidence exists of their contributions to creating resilient, digitally enabled ecosystems that foster interconnectedness, inclusivity, and innovation uptake by Industry and society?

Methodological Approach The research applies a mixed-methods approach that integrates qualitative and quantitative data sources, combining policy analysis, empirical reviews of programme data, and in-depth case studies. A multi-level analytical framework is used to assess contributions to the digital transition across four levels: 1. Meta-Policy Level: Examining strategic alignment of EU-level R&I and sectoral policies with digital transformation priorities, including their responsiveness to societal and industrial needs through strategic programming and alignment with Sustainable Development Goals (SDGs), new industrial policy and economic security for Europe. 2. Programme Design and Implementation: Analysing intervention portfolios at the thematic level, including funding allocations, directionality of research topics, and mechanisms to ensure coherence across funding instruments and innovation domains. 3. Project-Level Contributions: Evaluating project-level outputs and outcomes, including technological advancements, research-based policy recommendations, and societal impacts and long term aggregated effects of the Programmes. 4. Knowledge Ecosystems: Assessing the creation and reinforcement of digitally capable ecosystems that connect diverse actors—industry, academia, SMEs, and civil society—to support innovation diffusion and market adoption.

Evidence is drawn from policy documents, programme evaluations, stakeholder interviews, and project data. Key analytical dimensions include: • Policy Directionality: Review of Strategic Research and Innovation Agendas (SRIAs), work programmes, and project calls to assess alignment with digital and industrial transformation objectives, such as improving competitiveness, driving systemic innovation, and supporting technology adoption. • Stakeholder Diversity: Evaluation of co-creation practices, inclusivity in programme design, and geographic and sectoral diversity among participants to identify barriers and enablers of broader engagement. • Systems Integration: Identification of mechanisms for fostering cross-sectoral collaboration, integrating value chains, and supporting transdisciplinary approaches in research and innovation.

Findings and Results Preliminary findings highlight significant progress in aligning the Framework Programmes with the digital transformation objectives of industries. Evidence includes: • Strategic Targeting and Directionality: Programmes demonstrate a strong focus on accelerating digital transformation through targeted calls and thematic priorities. Key focus areas include artificial intelligence, advanced manufacturing, quantum technologies, and circular economy-driven digital solutions. These interventions align with broader EU strategies for digital and industrial transition goals, contributing to global competitiveness and sustainable growth. • High-Impact Innovation and Adoption: The Framework Programmes support a mix of high-risk, high-impact research initiatives and applied projects. Outputs include prototypes, pilot systems, and testbeds for disruptive technologies. Early results indicate improved industry uptake of innovations, though challenges remain in bridging technology readiness levels (TRLs) and scaling early-stage technologies to market. • Cross-Sectoral Integration: While progress has been made in fostering collaboration across sectors, further efforts are needed to address persistent silos and enhance cross-sectoral knowledge diffusion. Specific initiatives, such as co-programmed partnerships, have been instrumental in strengthening coordination across industries and research fields. • Inclusivity and Stakeholder Engagement: Initiatives to engage diverse stakeholders—SMEs, start-ups, civil society, and underrepresented regions—are yielding promising results. The use of Financial Support to Third Parties (FSTP) schemes and co-creation mechanisms has enabled broader participation, particularly in localised and place-based innovation projects. • Knowledge Ecosystem Development: The programmes are beginning to establish interconnected ecosystems that link industry, academia, and policymakers. These ecosystems foster resilience, agility, and sustainability by supporting knowledge diffusion, talent circulation, and collaborative problem-solving.

Despite these advances, several challenges persist: 1. Operational Synergies: Greater coherence between EU-level programmes, such as Digital Europe, Connecting Europe Facility, and Horizon Europe, is needed to optimise resource allocation and impact. Current planning cycles and funding mechanisms present alignment challenges. 2. Innovation Scaling: While disruptive innovations are supported, mechanisms to ensure scaling and market adoption remain underdeveloped. Improved alignment of funding instruments across TRLs is necessary to bridge the gap between early-stage research and commercial application. 3. Standards and Regulation: Digital transformation requires enhanced integration of standards and regulatory frameworks. Engagement with standardisation bodies remains limited, and stronger mechanisms are needed to ensure that R&I outcomes inform policy and regulatory development.

Conclusion This study highlights the critical role of EU R&I Framework Programmes in driving the digital transformation of industries. By embedding targeted digital transition objectives into policy frameworks, programmes foster systemic innovation and enhance the competitiveness of European industries. However, the findings also show areas for improvement, including better integration of research and market adoption pathways, stronger coherence across funding instruments, and enhanced alignment with international standards and regulations.

Future research and policy development should focus on scaling innovations, strengthening cross-sectoral collaboration, and leveraging knowledge ecosystems to address complex global challenges. The insights from this study provide valuable lessons for shaping the next generation of R&I programmes, ensuring that they contribute effectively to Europe’s digital and industrial transition, laying the ground for other deployment programmes such as the Digital Europe Programme.

ABSTRACT. Research Questions In this paper, we study the question of whether government support of business enterprise R&D (aka ‘BERD’) affects innovation output qualitatively. The theoretical background is the extensive empirical literature that has evaluated the effectiveness of fiscal R&D incentives. The findings of this longstanding policy-oriented work indicate that public BERD support, broadly speaking, has a positive effect. Bloom et al. (2019:180) conclude that RD&I tax credits and direct public funding ‘seem the most effective measures in innovation policy toolkit’. Much of the work has focused on the return to the public investment to spur private sector R&D investment. This ‘input additionality’ literature indicates that the effect is positive and sizeable, depending on specific settings (eg Almus and Czarnitzki, 2003). This paper contributes to the work that has tried to evaluate the effect on the firm’s innovation output itself (eg Cappelen et al., 2012). The question of ‘output additionality’ matters if the policy goal is to affect not only the level of innovation in the economy, but also the direction of that innovation, for example to promote more climate-friendly solutions. In this context, the policy background is important too. Government spending on BERD support is large, accounting for roughly 460 billion dollars or 0.21 of GDP in OECD countries according to the latest available numbers (2021, see OECD 2023). Spending has also increased substantially (%50 percent since 2000) and there’s been actively rebalanced between different types of measures. As the level of support is actively adapted and rebalanced between direct and indirect policies, it is important to continue to improve our understanding of the effects that the full range of R&D grants and tax credits is having on innovation in the private sector (OECD, 2021). The proposed paper contributes policy-relevant evidence about the effect that BERD support has on innovation output. In general, this is an area where further research is needed. Broadly, the relatively small number of studies have found that BERD support has positive effects on innovation (eg Dechezleprêtre et al. 2016). However, results vary and tend to be sensitive to a range of study-specific factors (Dimos & Pugh, 2016). The immediate background for this paper, however, is recent work that took a ‘completest’ approach to the question. In two recent studies (Nilsen et al, 2020; Iversen et al, 2024) we find that the effect of BERD support tends to diverge: for firms that have not been R&D active, we find a strong and positive effect along the extensive margin; for incumbent R&D active firms, the effect on the intensive margin is marginal. This divergent effect helps to explain why results from the output additionality work are not clearer. At the same time, it begs the question of whether there is another effect among the important incumbent R&D active firms that we are missing by tallying not only patents but patents together with other IPRs. This is what we study in this paper. Data and Methodology This paper builds on a string of empirical work that has followed the well-documented Norwegian case. Following Iversen et al (2024), we follow the ‘new economics of industrial policy’ by emphasizing improved measurement, careful application of causal inference, and a more ‘nuanced and contextual understanding of the effects of industrial policy (Juhasz et al, 2023). Data integration: Three main data sources yield a comprehensive panel dataset: (i) The Business registry consisting of firm-level data for the full population of active Norwegian limited liability business enterprises. (ii) The database on public BERD support consisting of detailed information about R&D subsidies from all major programs whether direct subsidies (grants) or tax credits.(iii) IPR applications (patents, trademarks, industrial design). These micro-data are linked at the enterprise-level to allow us to study the relation between BERD support and IPR outcomes over a period extending at least from 2002 (when the Norwegian R&D tax credit program started) until 2021. Key dimensions of the underlying industrial organization are traced at the micro-level over two decades. The panel observes firms reporting formal R&D and/or innovation activity, and which firms receive what form of R&D subsidy over time. Measurement: The target literature has proxied innovation outcomes using measures from R&D/Innovation Surveys, from firm-linked patenting (e.g. Bronzini & Piselli,2016) or from both (Cappelen et al. (2012). Our innovation metric is based on a composite dataset for the three types of IPR. This goes beyond a simple patent count approach. But an approach based on IPR counts more generally still gloss over the link between the IPR and the specific R&D policy instruments (eg green patents for direct BERD support on climate mitigation). The paper introduces qualitative aspects of firm level innovation outcomes to address such issues. How then to proxy the quality of innovation output? We derive measures about patent quality including application channel, patent family size, renewal, as well as bundles with trademarks and industrial designs. In this sense, patent quality is a function of a potentially multifaceted trajectory of events following the initial patent filing, leading to methodological questions. Method: Patent quality is framed in terms of an event study design. Efforts are made to mitigate challenges this poses to identification of causal effects in the presence of time-varying confounding factors. To address such challenges, we will draw upon ongoing research in this area (see Iversen at al., forthcoming 2024) that apply causal machine learning methods (Belloni, 2016; Chernozhukoc et al, 2018) to identify and estimate treatment effects in non-linear, dynamic models. Findings Our preliminary findings build on Nilsen et al (2020) and Iversen et al, (2024) that fiscal stimulus tends to have greatest impact on previously non-innovative firms. The impact of support measures, alone or in combination, is on the extensive rather than intensive margin. Findings from the extension into the quality of IPR outcomes is ongoing and have not yet finalized.

ABSTRACT. The proposed paper addresses one of the questions intended to animate presentations and discussions in the Conference: to what extent have science and innovation policy interventions generated economic and social impact? Or from a slightly different perspective: what prevents science and innovation policy interventions to generate economic and social impact?

Drawing from the Brazilian case, we start showing compelling evidence that policies promoting innovation have not been effective. Since the late 1990s science and technology policy in Brazil has established innovation as a priority. Starting with the creation of the Sectoral Funds, in 1999, followed by a number of subvention schemes and fiscal incentives to promote business R&D, such as Green and Yellow Fund (2001), Innovation Law (2004), Lei do Bem (2005) and so on, there was a significant increase in the number of instruments aiming to further innovation. Despite all of these initiatives, the results have been disappointing, as there is no indication that Brazilian companies are improving their innovation performance.

We argue that so far policies have been too much concentrated on the supply side, on measures to induce firms to invest in R&D. The demand side has been neglected; the measures that make firms want to spend on R&D and seek to innovate.

We show that globalization changed the conditions under which the Brazilian economy functioned therefore limiting the effectiveness of growth strategies followed by the country until then. A key feature of globalization, which is the functional integration of geographically dispersed activities of multinational companies, has implications for the economic dimension of the domestic markets of the countries in which these companies operate. As a result, development strategies based on the size of the domestic market became highly compromised. This take is valid not only for Brazil but for Latin America overall.

Failing to adapt to this new scenario brought up by globalization implies that firms have limited stimulus to engage in innovation activities. To be effective, policies must focus on both the supply and demand side. That is, it is necessary to provide physical and human capital necessary to carry out innovation activities, institutions that facilitate innovation, subvention schemes, fiscal incentives, governmental programs that promote innovation, and so on. Nevertheless, it is equally necessary to create an economic environment, particularly a competitive regime, which compels firms to innovate.

ABSTRACT. Corporate research laboratories of the 20th century advanced science and technology enormously. The iconic Bell Labs alone won eleven Nobel prizes and four Turing awards and launched the fields of microelectronics and information science. Some such laboratories still exist today, but with a much-diminished role in the national research ecosystem [1]. In this talk, building on our recent modern synthesis of the nature and nurture of research [2], we distill the reasons such laboratories were so productive into a set of funding and organizing principles for how they might be re-established in a 21st century form. Our aim is for these principles to catalyze debate on policy opportunities, which we believe could be transformative for national research productivity, and on policy challenges of practical implementation.

A first principle is “inspire, but don’t constrain, research by particular use.” As is well known by corporate scientists and engineers, intimate familiarity of use—human-desired interactions with the real world—stimulates a rich space of problems associated with such use. As is also well known by corporate scientists and engineers, however, the immediacy of solving exactly those problems, of constraining research to particular use, is typically paramount. Occasionally (often, even, if one is intentionally looking), however, surprise intervenes: an unexpected new idea for solving the problem more generally but less immediately; or an unexpected new observation that contradicts expectation [3]. Pivoting to that surprise means moving into adjacent problem spaces—a detour that likely doesn’t solve, at least not very efficiently, the original problem. But pivoting to that surprise is the essence of research, of disruption to existing knowledge [4], and a signature of Bell Labs 1.0 research. In other words, corporations, with their problem-rich use environments, are powerful starting points for research, but, because of competitive market pressures, are inherently unable to support the pivoting to surprise. Our proposed Bell Labs 2.0 policy solution for this is a public-private partnership in which corporations host and cost-share, but funders with public-goods benefit in mind provide majority funding, for such research. Such a public-private partnership would be different from academic research in being embedded in real-world problem-rich environments, and from both traditional corporate “development” and the newer focused research organizations (FROs) in seeking surprise.

A second principle is “fund and execute research at the institutional, not individual researcher, level.” As is well known, virtually all goods and services in modern economies are produced by firms, not by individual contractors who manage their own individual transactions with a complex web of customers and suppliers. The production of goods and services is much more optimally orchestrated by empowered leaders with tacit contextual knowledge of complex resources, including that most complex of resources, people [5]. Most importantly for our purpose, new knowledge, particularly the surprising and disruptive new knowledge associated with research, is no different—its production is likewise more optimally orchestrated, as it was at Bell Labs 1.0, by empowered leaders. Our proposed Bell Labs 2.0 policy solution is thus a “firm” organizational model—one in which a research institution is block funded and whose leaders actively orchestrate and nurture researchers as employees, not contractors. Such a “firm” organizational model for research is opposite to the matrix-funded models that dominate the current research landscape—models in which researchers are employed by other institutions but contracted by funding organizations to perform research.

A third principle is “evolve research institutions by retrospective, competitive reselection.” As is well known, selection and reselection of leadership is an unsolved problem in organizational management. Great leaders are not guaranteed and, because in a “firm” organizational model leaders are of extreme importance, great research institutions are also not guaranteed. Moreover, there is much we do not know about the optimal organization of a research institute and how that optimal organization might vary with the advance of technologies for learning and collaboration [6]. Our proposed Bell Labs 2.0 policy solution is thus a network of Bell Labs 2.0’s: each hosted and cost-shared by a different corporate host; each majority funded by an organization that has the broader public good in mind; and, most importantly for this purpose, all competing with each other for re-selection. Such an internally competitive network of Bell Labs 2.0 is different from most existing research institutes which are captive to their research funder. For example, Janelia Research Campus is captive to the Howard Hughes Medical Institute and is virtually guaranteed funding without competition.

In summary, we believe the policy opportunity for these “Bell Labs 2.0” research constructs is enormous—the 2024 Chemistry Nobel Prize is a modern testament to research that can be done at corporate research labs, if properly configured. There are also policy challenges, of course, including: how can “inspire, but don’t constrain, research by particular use” be best operationalized; how might differences between the production of goods and services vs of knowledge alter trade-offs between firm vs contractor research models; and how can competition between research institutes in potentially vastly different knowledge domains be operationalized? We hope this talk catalyzes debate on these policy challenges, as well as on the policy opportunities.

[1] A. Arora, S. Belenzon, L. Sheer (2021). “Knowledge spillovers and corporate investment in scientific research.” American Economic Review 111, 871-898. [2] V. Narayanamurti, J.Y. Tsao (2024). How technoscientific knowledge advances: A Bell-Labs-inspired architecture. Research Policy 53, 104983. [3] A Bell Labs 1.0 example of the first kind of surprise was Claude Shannon’s information theory, a general way of thinking about efficient information communication. A Bell Labs 1.0 example of the second kind of surprise was Karl Jansky’s discovery of radio signals from the Milky Way, leading to his founding of the field of radio astronomy. [4] M. Park, E. Leahey, R.J. Funk (2023). “Papers and patents are becoming less disruptive over time.” Nature 613, 138-144. [5] O.E. Williamson (2010). “Transaction cost economics: The natural progression.” American Economic Review 100, 673-690. [6] M.A. Nielsen, K.J. Qiu (2022). “A Vision of Metascience: An Engine of Improvement for the Social Processes of Science.”

ABSTRACT. Background The commercialization of university research has become increasingly crucial for technological innovation and economic development. Following the success of the Bayh-Dole Act in the United States, many countries have adopted similar IP rights reforms to enhance university technology transfer. In China, despite implementing comparable IP rights policies since 2000 that grant universities patent rights and establish revenue-sharing mechanisms, the commercialization rate of university patents remains disappointingly low at below 5%. This limited success presents an intriguing puzzle: why do similar IP rights reforms yield dramatically different outcomes across institutions? The challenge is particularly complex in China's institutional context, where university-generated intellectual property has traditionally been classified as state-owned assets, subjecting technology transfer decisions to rigorous government oversight and auditing requirements. This institutional arrangement creates significant risks for universities in asset valuation and transfer pricing, potentially undermining the intended effects of IP rights reform. The variation in technology transfer performance across universities, despite facing similar institutional constraints, suggests that the effectiveness of IP rights reform may depend on its interaction with other organizational and environmental factors. Understanding these complex relationships is crucial for improving the design and implementation of technology transfer policies. Research Questions Given this context, our study addresses three interrelated research questions: First, why do similar IP rights reforms yield dramatically different technology transfer outcomes across universities operating under the same national policy framework? Second, what configurations of institutional arrangements, organizational capabilities, and environmental conditions enable successful university technology transfer? Third, how does the effectiveness of IP rights reform implementation depend on its interaction with universities' internal capabilities (scientific discovery, technological application, human capital, and reputation) and external conditions (regional economic vitality)? These questions move beyond the traditional focus on policy adoption to examine the complex conditions under which IP rights reforms actually deliver their intended outcomes. Research Methods To unravel this puzzle, we employ fuzzy-set Qualitative Comparative Analysis (fsQCA) to examine 28 cases from China's prestigious 985 Project universities. Our research design offers two key methodological innovations. First, we measure technology transfer performance through actual contract values rather than conventional patent-based metrics, thereby capturing the true extent of commercialization success while avoiding the limitations associated with patent counting and valuation ambiguities. Second, we adopt a configurational approach to analyze how IP rights reform interacts with multiple conditions to produce successful outcomes. These conditions include scientific discovery capability (measured by research output and quality), technological application capacity (indicated by applied research and development activities), human capital (represented by research staff qualifications and industry experience), university reputation (based on national and international rankings), and regional economic vitality (reflected in industrial development and market demand for innovation). This comprehensive analytical framework allows us to identify complex combinations of conditions that enable or constrain successful technology transfer. Findings Our analysis reveals several important patterns. First, the necessity analysis shows that no single condition, including IP rights reform implementation, is sufficient alone to guarantee high technology transfer performance. However, the persistence of traditional state-owned asset management approaches (without effective IP rights reform implementation) emerges as a necessary condition for poor performance, highlighting the fundamental importance of institutional arrangements. Second, our sufficiency analysis identifies three distinct pathways to successful technology transfer: (1) effective implementation of IP rights reform combined with strong accumulated knowledge capital, enabling universities to leverage their existing research capabilities; (2) reform implementation paired with robust research capacity and complemented by either strong human capital, established university reputation, or favorable economic conditions; and (3) reform implementation coupled with strong regional economic vitality, which can compensate for relatively modest university capabilities. These findings suggest that while effective IP rights reform implementation is crucial, its success depends on specific organizational and environmental configurations. Contribution This study makes several significant contributions to our understanding of university technology transfer. First, we advance the methodological approach by directly analyzing technology transfer contract values, addressing fundamental limitations in previous studies that rely solely on patent metrics. This innovation provides a more accurate picture of actual commercialization success and captures various forms of technology transfer that might not be reflected in patent statistics. Second, we enrich theoretical understanding of research management and science & technology policy by demonstrating that the effectiveness of IP rights reform depends on complex configurations of institutional, organizational, and environmental factors. This finding challenges the simple policy transfer approach that assumes similar reforms will yield similar results across different contexts. Third, our identification of multiple pathways to successful technology transfer provides valuable insights for policymakers and university administrators. It suggests that successful implementation of IP rights reform requires careful consideration of institutional contexts and complementary capabilities. Universities may need to develop different strategies based on their specific strengths and environmental conditions to achieve effective technology transfer. Finally, our research contributes to the broader literature on institutional reform by showing how the same policy instrument can yield varying outcomes depending on its interaction with other organizational and environmental factors. These insights are particularly valuable for emerging economies seeking to enhance their innovation capabilities through institutional reforms.

ABSTRACT. Research Question University incubators can play an important role in contributing to the innovation system by enabling technology and knowledge spin-offs from university to society (Fischer et al. 2021). Universities may set up incubators and accelerators in order to enable faculty, students, graduates and external entrepreneurs to enable technology spin-offs, to start and grow ventures that may contribute both to the local economy as well as to create a revenue stream for the university (Kolympiris & Klein et al 2017; Mian 1996).

In India, Higher education institutions look to take new technologies and ideas out of labs and into the market primarily through the route of venture creation via campus-based incubators. The invention to innovation trajectory is therefore squarely routed in the idea that Schumpeter’s entrepreneur will enable innovation and commercialisation of technology. Incubators have, therefore, mushroomed over the past decade across University campuses in India (Thillairajan & Jain, 2013; Sonne, 2014). However, we have limited information about how and, if, incubators are able to enable this lab to market trajectory in India (Vaishnav et al, 2024; Narayanan & Shin, 2019; Loganathan & Subrahmanya, 2021; Jansen et al. 2015). This paper contributes to filling this gap by studying how university incubators have evolved over the last decade.

McAdam et al. (2015), consider the context, relationships and stakeholders within the incubation ecosystem that enables the incubator to function. Here the incubator is influenced by: 1) the wider macro environment of the region and country; 2) The meso environment of the higher education institution – the organisational context including the type and focus of the institution, the quality of expertise through faculty, the culture of the institution including the relations with students and alumni, and the administrative relationship with the institution; 3) the micro environment within the incubator- its processes and areas of focus.

Drawing on McAdam et al. (2015), this paper studies the evolution of three higher education incubators and their respective ecosystems in India over the past decade. We ask: 1) How do these incubators support the transformation from lab to market via venture creation?; 2) What does the incubator’s ecosystem look like and how does it influence its ability to operate?; 3) How have the incubators evolved over the past decade?

Methodology This study, which is a work in progress, uses secondary archival data, as well as primary research in the form of semi-structured interviews during 2017 and late 2024/early 2025. A qualitative approach entailing multiple case studies was selected in order to allow for the study of complex issues. Using multiple cases makes it possible to study patterns, similarities and differences across cases while reducing the chance of coincidental occurrences (Eisenhardt, 1991; Yin, 2003). Purposeful sampling was used to select incubators representing a range of different factors, including type of higher education institute host, geographic location and sector focus. Semi-structured interviews were held November-December, 2017 with representatives of senior management from three incubators. Interviews were approximately one-two hours long per incubator, over the phone or in person. Interview notes were coded for similarities and differences. Based on interview notes, a priori themes were noted across different domains and information was then tabulated from different interviews to create a set of themes. In 2024, we use the first set of interviews and analysis to create a set of questions for follow-up interviews that will take place late 2024 and early 2025. The follow-up interviews will include four to five interviews (including senior managers, entrepreneurs, and mentors) for incubator to enable a deeper understanding of each incubator. The follow—up interviews will be coded and analysed using conventional qualitative methods tools. The analysis will follow a grounded and iterative process (Ritchie and Lewis, 2003), and will be strengthened through triangulation (Yin, 2003) of secondary archival data, case study interviews and review commentary.

Preliminary Findings Based on the original interviews in 2017, and secondary research during 2024, we found that incubators offer broadly similar types of support (infrastructure including office space, access to labs; non-financial support such as mentoring and services such as access to legal and financial advice; financial support; fundraising support, networks). Operationally, there is a need to experiment and test different models, and for the incubator to be entrepreneurial in order to offer services that startups are keen on.

With respect to the ecosystem, what makes a key difference, is the team that runs the incubator and the ecosystem of stakeholders providing support within and outside the university. Incubators across the board found being hosted at a higher education institution beneficial both for the brand and access to key stakeholders including government, but also for the access to a potential pipeline of student and faculty ventures and new technology, as well as the access to great infrastructure and to talent. Host institutions in some cases are also important sources of funding. At the core of successful incubators at higher education institutions is buy-in from senior leadership at the host institute. While it is clear that higher education institutions provide a lot of benefits, as hosts of incubators, it is important to have considerable buy-in from senior leadership at the host institute. Several incubators stressed the need to be able to operate independently. Lastly, building the ecosystem and collaborating with other stakeholders in the ecosystem matters in order to provide support for venture creation. The findings will be significantly updated after the follow-up interviews are completed.

The paper will contribute to a considerable gap in research and knowledge about the functioning of incubators based at university incubators in India.

ABSTRACT. 1. Introduction

Universities drive innovation through intellectual property, technology transfer, and research collaboration, essential to national and regional innovation systems (Nugent & Chan, 2023; Perkmann et al., 2013). The shift to a knowledge economy has expanded universities’ missions to include fostering economic and social development (Compagnucci & Spigarelli, 2020).

This study examines how internal characteristics of Brazilian federal universities—such as size, governance, and scientific output—affect their innovation activities. Using data from all 69 federal universities, we analyze factors influencing patenting and research collaborations. Our findings highlight key institutional traits linked to innovation, providing insights into how universities contribute to Brazil’s knowledge economy. The article covers literature, methodology, results, and conclusions.

2. Literature Review

Universities are increasingly integral to innovation systems, translating research into commercial applications through patenting, technology transfer, and industry partnerships (Aghion & Howitt, 2008; Mansfield & Lee, 1996). This shift is aligned with the Third Mission (TM), which emphasizes universities' contributions to economic and social development, and with the growing proximity of research outputs to outcomes and innovation (Compagnucci & Spigarelli, 2020).

Factors influencing university patenting include its size, the institutional support, and fields of research, with larger and older institutions often patenting more, particularly in engineering and chemistry (Fisch et al., 2015; Yamaguchi et al., 2019). Additionally, patent quality, as seen in Brazilian universities, is significant, as impactful patents contribute more effectively to innovation ecosystems (Méndez-Morales et al., 2022).

The literature also suggests a positive correlation between publishing and patenting, challenging the notion that commercialization detracts from research output (Stephan et al., 2007). Instead, patenting often complements scientific productivity, particularly in knowledge-intensive fields (Breschi et al., 2008). However, successful academic patenting requires supportive institutional frameworks, as highlighted in studies on emerging economies (Murat AR et al., 2023).

3. Method

This exploratory and descriptive study uses quantitative data from Brazil's Federal Court of Auditors (CGU) audit project "Economics of Innovation in Federal Public Universities", and bibliometric data extracted from large data sources. The audit examined how strategies within Brazil's innovation ecosystem contribute to innovation-driven research, considering the Triple Helix model. Complementary bibliometric data allowed to correlate primary data collected by CGU with secondary data collected by the authors.

3.1 Research Question and Variables

The study aims to determine:

Research Question: a) To what extent are the constitutive characteristics of Brazilian federal universities correlated with their performance variables? b) What are the factors that most influence technological production and transfer?

Variables Used:

Constitutive Variables: University size (number of faculty), age, and geographic location. Performance Variables: Scientific publications, citations weighted per area of knowledge and timeframe, intellectual property (IP), patent citating scholarly outputs, articles co-authored with companies, technology transfer rates, and internal regulatory guidelines towards IP and tech-transfer.

3.2 Data Collection

Data were gathered through:

Census Surveys: Questionnaires covering intellectual property, technology transfers, and institutional strategies related to innovation. Secondary Data: Information from the Ministry of Education, Sucupira Platform (from the Brazilian Federal Agency for Support and Evaluation of Graduate Education - CAPES), Scopus database, and Scival platform.

3.4 Data Analysis

Bibliometric analysis of publications (2011–2022) was performed using Scopus and Scival. Key indicators related to technology transfer, intellectual property and institutional regulations were selected. Pearson and Spearman correlation analyses and Qualitative Comparative Analysis (QCA) were conducted.

4. Empirical Context

Brazil’s higher education system includes over 2,500 institutions, with about 87% private and 13% public. Despite their smaller share, public universities, particularly federal ones, are central to scientific research and PhD training, employing more than half of Brazil’s PhD faculty and contributing significantly to scientific output (INEP, 2022). In 2023, Brazil ranked 13th globally in scientific production, largely due to federal universities, which excel in health sciences, agrarian sciences, and engineering (Clarivate, 2024).

Federal universities, under the Ministry of Education (MEC), have standardized governance, with deans appointed by the President of Brazil following academic consultations. These institutions vary widely in size, age, and regional distribution, reflecting Brazil’s economic and infrastructural diversity. For example, some are over a century old, while others were established within the last 20 years, highlighting the range in institutional characteristics.

This study is the first comprehensive analysis covering a large set of indicators of scientific, technological and institutional regulation of all Brazilian Federal universities.

5. Results

Here are some findings yet to be discussed and developed: • There is no correlation between constitutive characteristics of universities (age and tradition, location and size) and their performance on variables of outputs. • The are strong correlations between scientific indicators (as for scholarly outputs, citations per field and timeframe and co-authorship with companies) and IP and tech-transfer. • The presence of internal regulations about IP and tech-transfer act as a factor that positively influences the rates of IP and tech-transfer. • There is evidence that tradition (in terms of age, size and location in more developed states) does not serve as a differentiating factor for universities being more oriented toward IP and tech-transfer.

ABSTRACT. Proof of Concept Centers (PoCCs) are an increasingly prevalent strategy for university technology translation (Bradley et al., 2013). PoCCs support technology transfer and commercialization by providing academic innovators with training, mentorship, and non-dilutive funding to de-risk technologies and reduce information asymmetries that could create barriers to commercialization (Gulbranson & Audretsch, 2008). While PoCCs are associated with better translational outcomes (Hayter and Link, 2015), there is limited evidence or understanding of how PoCCs impact commercialization trajectories across a wide range of technologies and for PIs with different backgrounds (Munari and Toschi, 2021; Battaglia et al., 2021; Abudu et al. 2022).

This paper asks how projects supported by PoCCs with different principal investigators (PIs) and technology characteristics vary in reaching three intermediate commercialization outcomes: patents filed, licensing agreements, and Small Business Innovation Research (SBIR) or Small Business Technology Transfer (STTR) award applications by startups founded around PoCC-supported technologies. These three outcomes are important to the technology transfer process and are of interest to the research and policy communities (Link, 2024). Our analysis shows that initial paths to outcomes within PoCCs vary based on technology type and the PI’s academic title, gender, and commercialization experience. These findings make novel contributions to the literature on PoCCs and technology transfer.

We use a dataset of 275 biomedical technologies that have received programmatic funding from 6 distinct PoCCs supported by the National Institute of Health (NIH): three centers funded under the NIH Centers for Accelerated Innovations (NCAI) beginning in 2014, and 3 hubs from the first iteration of the Research Evaluation and Commercialization Hubs (REACH) program beginning in 2015. For each PoCC-supported technology, our dataset tracks the trajectory of the technology over a period of 4 to 9 years, observing outcomes associated with the core technology supported.

The dataset is unique in the timespan of observations following funding, the scope across multiple PoCCs, and the ongoing efforts to update and ensure validity of outcomes. These data are summarized by Anderson et al. (2022) in a programmatic review but have not previously been used in an empirical analysis. While Munari & Toschi (2021) tracks outcomes for multiple European PoCCs, data of this type have not been previously used in the United States.

To estimate the impacts of project-level characteristics on patents filed, licensing agreements, and SBIR/STTR awards, we employ a Tobit estimation to predict the number of outcome events. We control for NIH PoCC program funding, technology type, and the PI’s title, gender, and commercialization experience. Controls for the six PoCCs in the data are included and standard errors are clustered at the center level. For robustness, models are also run with ordinary least squares (OLS) and Probit estimations, which support initial fundings.

Results show important variation in outcomes achieved by PoCC-supported projects based on PI characteristics and technology type. PIs with prior commercialization experience had higher numbers of SBIR/STTR applications and were much more likely to pursue licensing agreements but had no significant differences in the number of patents filed. Full professors more likely to license technologies, while assistant professors were more likely to file patents or apply for SBIR/STTR awards. Associate professors had low rates across all outcomes. Female PIs are estimated to be more likely to patent and apply for SBIR/STTRs but had fewer licensing agreements. However, we found no statistically significant differences between male and female PIs. Among technology categories, drug projects had fewer outcomes than other technology types, likely due to lengthy timelines required for development and regulatory approval. Medical devices and diagnostics had the highest patenting rates. We also observe a strong positive correlation between the level of NIH funding and the frequency of licensing agreements, with a weaker yet positive association with patents filed and SBIR/STTR applications.

The experiences of projects in NIH PoCC programs support show the importance of considering different translational pathways and trajectories which may be relevant to different projects. They also support the efficacy of PoCC funding in spurring academic research with commercialization potential towards translation activities. Policymakers can use results to better support a range of innovators and nascent technologies.

Works Cited

Abudu, R., Oliver, K. and Boaz, A. (2022) ‘What funders are doing to assess the impact of their investments in health and biomedical research’, Health Research Policy and Systems, 20(1). https://doi.org/10.1186/s12961-022-00888-1.

Anderson, B. J., Leonchuk, O., O’Connor, A. C., Shaw, B. K., & Walsh, A. C. (2022). Insights from the evaluations of the NIH Centers for Accelerated Innovation and Research Evaluation and Commercialization Hubs programs. Journal of Clinical and Translational Science, 6(1), p. e7. https://doi.org/10.1017/cts.2021.878.

Battaglia, D., Paolucci, E., & Ughetto, E. (2021). The role of Proof-of-Concept programs in facilitating the commercialization of research-based inventions. Research Policy, 50(6), 104268. https://doi.org/10.1016/j.respol.2021.104268

Bradley, S.R., Hayter, C.S. and Link, A.N. (2013). Proof of Concept Centers in the United States: an exploratory look. The Journal of Technology Transfer, 38(4), p. 349–381. https://doi.org/10.1007/s10961-013-9309-8.

Gulbranson, C.A. and Audretsch, D.B. (2008). Proof of concept centers: accelerating the commercialization of university innovation. The Journal of Technology Transfer, 33(3), p. 249–258. https://doi.org/10.1007/s10961-008-9086-y.

Hayter, C.S. and Link, A.N. (2015) On the economic impact of university proof of concept centers. The Journal of Technology Transfer, 40(1), p. 178–183. https://doi.org/10.1007/s10961-014-9369-4.

Link, A.N. (2024). Public Sector Technology Transfer. Edward Elgar Publishing. Available at: https://www.e-elgar.com/shop/usd/public-sector-technology-transfer-9781035310524.html

Munari, F. and Toschi, L. (2021). The impact of public funding on science valorisation: an analysis of the ERC Proof-of-Concept Programme. Research Policy, 50(6), p. 104211. https://doi.org/10.1016/j.respol.2021.104211.

ABSTRACT. Since the enactment of the Bayh-Dole act, universities have increasingly become a source of innovation that leads to commercial products. In the U.S. alone, university researchers publish nearly one million articles annually. Only a small percentage of the technologies described by these articles are formally disclosed by universities through their technology transfer offices; even fewer are patented and/or licensed for commercial application. For example, while researchers at the University of Michigan publish well over 10,000 papers annually, less than 5% of these publications result in technical disclosures, with even fewer advancing to patent applications. In addition, many tools and other byproducts of research have commercial utility but are not reported to university technology transfer offices.

Our preliminary research, based on analysis of more than 80,000 research articles collected over an 18-month period, suggests that the number of university publications with high commercial potential significantly exceeds current disclosure rates. We developed and validated two complementary AI-based approaches to identify commercially promising research outputs: one focused on identifying disclosure-ready material, and another designed to replicate the evaluation criteria used by technology transfer professionals.

In validation experiments with licensing experts from the University of Michigan and the University of Chicago, our system demonstrated remarkable effectiveness. When presented with 100 previously unexamined research articles (50 classified by our model as commercially promising), the blinded expert reviewers identified 39% and 51% as promising candidates for commercialization, respectively. More significantly, when they were then provided with the AI-generated reasoning for each classification, the experts independently modified some of their responses, increasing their alignment with the AI and each other. As a result, the inter-rater reliability between experts improved dramatically from 66% to 82%, suggesting that our system not only identifies promising research but also provides useful context for human evaluators.

Our preliminary work already has important implications for science policy and university technology transfer: 1. There appears to be a substantial gap between potentially commercializable research and actual disclosures, suggesting that current manual screening processes may be missing valuable opportunities for economic impact.

2. AI-assisted screening could help TTOs more efficiently allocate their limited resources, allowing them to focus on the most promising opportunities while expanding their overall screening capacity.

3. More comprehensive identification of commercializable research could significantly increase universities' contributions to economic development, better fulfilling the intentions of the Bayh-Dole Act.

4. The high inter-rater reliability achieved with AI assistance suggests potential for more standardized evaluation processes across institutions.

This research suggests that by leveraging AI-assisted screening, universities could significantly increase their innovation pipeline while maintaining high-quality standards for commercial development. The next stage of our work focuses on incorporating actual licensing outcomes to refine our predictive capabilities, moving beyond disclosure worthiness (predicting expert judgments) to directly assess commercial viability (predicting executed licensing outcomes). The dramatic improvement in inter-rater reliability with AI assistance (66% to 82%) suggests that automated screening tools could not only increase the quantity of identified commercial opportunities but also improve the consistency and quality of evaluation processes. This could lead to more equitable access to commercialization opportunities across different research domains and institutions.

ABSTRACT. Introduction The availability of scientific literature is essential to scientific progress. Open Access (OA) refers to peer-reviewed publications that can be accessed online without requiring a journal subscription or other fees from readers (Piwowar et al., 2018). OA articles have been identified as important to scientific and technical progress in low- and middle-income countries and at institutions with constrained resources (Powell et al., 2020). This presentation will explore three facets of OA publications: 1) country-level outputs by OA classification, 2) scientific impact by OA classification at the country level, and 3) world maps displaying the proliferation of Gold OA (freely distributed online immediately by a journal) relative to closed access publications.

Methods The OA status of scientific publications was determined using the Scopus (Elsevier) bibliometric database and associated with countries by the affiliation of each coauthor. Open Access publications were classified into four categories: Gold OA, Green OA, Other journal-based OA, and closed access. Gold OA denotes journals that are entirely OA as a matter of policy. Green OA denotes articles that are self-archived by authors in independently open repositories (e.g. ArXiv). Other journal-based OA types include hybrid OA (journals with a mix of closed and open access articles) and bronze OA (journals that make articles conditionally available) which are combined for this analysis. Articles were categorized exclusively into the highest level of OA they appear (e.g., articles that are available both through Gold OA and Green OA are coded as Gold OA).

Measures of scientific impact were calculated from citations accumulated by each paper. The Highly Cited Articles (HCA) ratio, used in Science and Engineering Indicators (SEI), proxies scientific impact at the country level. For a given year, the HCA ratio is the percentage of a country’s total publications that appear in the top 1% of all publications (with a minimum two-year citation window). The top 1% is calculated separately for each scientific field to balance out citation differences among fields. This presentation will show HCA calculations for each of the OA types at the country level.

To illustrate the relative tendency for a country’s authors to publish Gold OA or closed access, we determined an index based on the ratio of Gold OA and closed access articles and generated world maps using this index for a series of years. For each country that publishes above the threshold of 30 fractionally counted articles, the number of Gold OA publications was divided by the number of closed access publications in a given year.

Results For each of the selected countries, the percent of Gold OA articles has increased between 2003 and 2022, and China and the European Union (EU) showed the greatest increases. The percent of closed access articles for the selected countries showed corresponding decreases over time, and by 2022, the EU exhibits the smallest proportion of closed access articles with 30.5%.

The HCA ratio was calculated at the country-OA level and is relative to a baseline of 1. Among selected top producers, there is a trend towards increased impact within the OA publication types and a reduced impact in closed access relative to the world. These trends show China increasing in all types. The United States had a dramatic decline in closed, a more modest decline in Gold and Green, and modest growth in other journal-based OA (hybrid/bronze). In 2020, the U.S. had lower HCA ratios than China in all categories other than Gold and the overall ratio with all categories pooled.

There are noticeable regional trends including high Gold OA publications in Africa and Southeast Asia. European nations are also exhibiting a relatively higher tendency to publish Gold OA, with several nations including France and the United Kingdom publishing a greater number of Gold OA articles than closed access. However, the United States and India still exhibit a greater prevalence of closed access articles.

Conclusion The rise of Open Access articles may indicate a change in the publication practices of researchers across the world, as OA becomes increasingly common over time. The United States and China , have demonstrated a growing tendency to publish OA over the past 20 years (National Science Board, 2023), even as their overall output is still predominantly closed access. In addition, many developing nations exhibited a higher tendency to publish Gold OA than closed access, suggesting a heightened importance of OA as a resource for freely sharing scientific findings to readers. However, with the higher article publication fees associated with OA publication, these shifting patterns are important to study from the perspective of global equity.

Works cited:

McKiernan, E. C., Bourne, P. E., Brown, C. T., Buck, S., Kenall, A., Lin, J., McDougall, D., Nosek, B. A., Ram, K., Soderberg, C. K., Spies, J. R., Thaney, K., Updegrove, A., Woo, K. H., & Yarkoni, T. (2016). How open science helps researchers succeed. eLife, 5, e16800. https://doi.org/10.7554/eLife.16800 National Science Board, National Science Foundation. 2023. Publications Output: U.S. Trends and International Comparisons. Science and Engineering Indicators 2024. NSB-2023-33. Alexandria, VA. Available at https://ncses.nsf.gov/pubs/nsb202333/. Nelson, A., Office of Science and Technology Policy (OSTP). (2022). Ensuring free, immediate, and equitable access to federally funded research. https://www.whitehouse.gov/wp-content/uploads/2022/08/08-2022-OSTP-Public-access-Memo.pdf Piwowar, H., Priem, J., Larivière, V., Alperin, J. P., Matthias, L., Norlander, B., Farley, A., West, J., & Haustein, S. (2018). The state of OA: A large-scale analysis of the prevalence and impact of Open Access articles. PeerJ, 6, e4375. https://doi.org/10.7717/peerj.4375 Powell, A., Johnson, R., & Herbert, R. (2020). Achieving an Equitable Transition to Open Access for Researchers in Lower and Middle-Income Countries [ICSR Perspectives]. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.3624782

ABSTRACT. The structure of the scientific communication system, influenced by research policies prioritizing "high impact" venues, plays a crucial role in shaping research agendas worldwide. These policies, aimed at promoting the publication of "excellent" research in mainstream indexing systems, often align with the interests of international readers rather than addressing local scientific and societal needs (López-Piñeiro & Hicks, 2015). This misalignment disproportionately affects the Global South, exacerbating national inequalities and undermining the critical roles played by local journals, such as training junior researchers, disseminating region-specific findings, and translating global scientific knowledge for local communities (Chavarro et al., 2017). While recent studies have explored the effects of research policies on locally oriented research (e.g., Mongeon et al., 2022; Orduña-Malea et al., 2024), a consensus on defining and operationalizing "local research" remains elusive.

Building on our earlier work (Di Césare & Robinson-Garcia, 2024), where we conceptualized local research through six complementary approaches, this study examines how local research dimensions are represented within two leading global indexing systems: Web of Science and Scopus. We focus on three dimensions of local research:

1. Locally Informed Research: Measured by the geographic concentration of references. 2. Locally Situated Research: Measured through the use of toponyms in journal content. 3. Locally Relevant Research: Measured by the geographic concentration of citations and authors.

To address these dimensions, we analyze a dataset of over 17,000 journals extracted from Dimensions, enriched with bibliometric indicators capturing these dimensions. Using fuzzy matching techniques, we merge this dataset with metadata from the Master Journal Lists and Journal Citation Reports from Web of Science and the Scimago Journal Rank list from Scopus, evaluating the inclusion and representation of local journals across regions, particularly in the Global South.

Our study investigates two interrelated research questions:

- How is local research from the Global South represented in Web of Science and Scopus? - Which scientific areas and disciplines are most affected by potential underrepresentation?

By highlighting the underrepresentation of local journals, this study contributes to a better understanding on the effects of research evaluation policies on national research priorities and topic delineation of scholars, as well as to the differences of cost in terms of citation impact and visibility for publishing local research between researchers from the Global North and the Global South.

REFERENCES Chavarro, D., Tang, P., & Ràfols, I. (2017). Why researchers publish in non-mainstream journals: Training, knowledge bridging, and gap filling. Research Policy, 46(9), 1666–1680. https://doi.org/10.1016/j.respol.2017.08.002

Di Césare, V., & Robinson-Garcia, N. (2024). What is local research? Towards a multidimensional framework linking theory and methods. Zenodo. https://doi.org/10.5281/zenodo.14033473

López Piñeiro, C., & Hicks, D. (2015). Reception of Spanish sociology by domestic and foreign audiences differs and has consequences for evaluation. Research Evaluation, 24(1), 78–89. https://doi.org/10.1093/reseval/rvu030

Mongeon, P., Paul-Hus, A., Henkel, M., & Larivière, V. (2022, September 7). On the impact of geo-contextualized and local research in the global North and South. Zenodo. https://doi.org/10.5281/zenodo.6956978

Orduña-Malea, E., Miguel, S., González, C. M., Arias, R. R., & Ortiz-Jaureguizar, E. (2024). An analytical framework for the study of geographical places in the scientific literature. Revista Española de Documentación Científica, 47(3), Article 3. https://doi.org/10.3989/redc.2024.3.1571

Rafols, I., Ciarli, T., & Chavarro, D. (2019). Under-reporting research relevant to local needs in the South: Database biases in rice research. In R. Arvanitis & D. O’Brien (Eds.), The Transformation of Research in the South Policies and outcomes. Éditions des archives contemporaines. https://digital.csic.es/handle/10261/226953

Ràfols, I., Molas-Gallart, J., Chavarro, D. A., & Robinson-Garcia, N. (2016). On the Dominance of Quantitative Evaluation in ‘Peripheral’ Countries: Auditing Research with Technologies of Distance (SSRN Scholarly Paper No. ID 2818335). Social Science Research Network. https://papers.ssrn.com/abstract=2818335

ABSTRACT. Generative AI (GenAI) technologies garnered significant attention with advancements such as ChatGPT in late 2022, profoundly influencing scientific practices. The unique capabilities of GenAI in tasks such as idea generation, literature review, data analysis, and academic writing have demonstrated a transformative impact on research activities, particularly on knowledge production behaviors across diverse research areas and regions. Despite its growing prominence, there is limited literature focus on the role of GenAI in shaping the research impact dynamics between the Global South and North, as well as between AI fundamental and application research.

AI technology impacts regions and research types differently due to disparities in resources, infrastructure, and priorities. Leveraging advanced technology and ample research funding, the Global North excels in AI-related research, particularly in AI fundamental research, having dominant innovation and global influence. In contrast, lacking of most advanced core technology, the Global South’s AI-related research tend to address localized challenges but with limited global reach. GenAI shapes research impact dynamics differently by amplifying resource and capability gaps, as well as lowering technical barriers and providing innovative tools. On the one hand, foundational advancements disproportionately benefit regions with better access, resulting in gaps in AI foundational research among regions. On the other hand, as a rapid, high-efficiency, and low-cost innovative tool, GenAI may increase the capability of conducting AI application research and addressing challenges for the Global South.

This study aims to empirically explore the impact of GenAI on research outcomes. Specifically, it examines whether the research impact (measured by journal impact factor percentile and journal quartile) of AI-related publications varies across regions (Global North vs. Global South) and research types (AI fundamental research vs. AI application research) before and after the emergence of GenAI. Additionally, it investigates the role of GenAI in bridging resource gaps while fostering equity, inclusiveness, and diversity. The study addresses these questions through two primary dimensions: (1) disparities between and within global regions, and (2) variations across research types. Based on these objectives, the study formulates the following hypotheses:

H1: The research impact of AI application research experienced more significant changes after the emergence of GenAI, because GenAI lowers the technical threshold, and GenAI's direct functions are more in line with the needs of applied research. H2: The increase in research impact in the Global North is more significant than in the Global South after the emergence of GenAI due to resource and technology advantages. H3: GenAI benefits AI application research in the Global South more, potentially bridging existing gaps. H4: GenAI benefits AI application research in the region that originally had lower scientific production, narrowing the gaps.

To address the research questions, this study utilizes nearly 500K AI-related articles from Web of Science Core Collections spanning 2021 to 2024, by retrieving AI-related keywords. We limited the data to the SCIE and SSCI database, and further narrowed the data to include only publications classified under the document type’s “article” and written in English.

The Global North-South classification follows UN Sustainable Goals Regions, using the corresponding author's affiliation as the primary indicator of regional classification and the first author's affiliation as a robustness check. The independent variable, ResearchType, classifies AI-related papers as either AI fundamental research or application research. This study defines papers involving computer science as AI fundamental research, and assigns the variable a value of 1; and defines papers that apply AI to other disciplines (such as economics, materials science, psychology, etc.) as AI application research. After sampling and manual inspection, there are a small number of papers that review or comment on AI research, and the sample proportion is negligible. Therefore, this study combines this part with AI application research. Since the GenAI model emerged at the end of 2022, the variable PostGAI is assigned with a value of 1 when the paper’s publication year is 2023 or 2024, and 0 if before 2023.

A triple-difference (DDD) framework with fixed effects is adopted in Eq.(1) to investigate the impact of GenAI on scientific practices and global research disparity. Additionally, event study is employed to estimate the effect of Generation AI across different time points and observe dynamic effects. Heterogeneity analysis and robustness checks are conducted for further analysis.

Using a triple difference approach combined with event study methods, this research systematically examines the dynamic effects of Generation AI on research impact disparities between the Global North and South, as well as between AI-related fundamental and application research. Our contributions to the literature are threefold. First, we advance interdisciplinary dialogue on technology diffusion and knowledge production by uncovering the academic ecological impacts of Generation AI. Second, we offer critical insights for policymaking and practical applications by quantifying regional and disciplinary disparities and analyzing their broader implications. Finally, we contribute to ongoing debates on technological equity and the future of global scientific practices, with a specific focus on GenAI's role in potentially exacerbating or mitigating existing inequalities.

ABSTRACT. Government tax policies are critical determinants of firms’ financial strategies, influencing their ability to acquire and manage capital. Prior research has extensively examined how tax policies such as corporate tax and R&D tax credits impact corporate behavior (Djankov et al., 2010; Hall & Van Reenen, 2000). Notably, R&D tax credits have been shown to enhance firms’ innovation activities by reducing the cost of research and development investments (Hall & Van Reenen, 2000; Czarnitzki et al., 2011). However, recent studies have highlighted that R&D tax credits disproportionately benefit large, profitable firms, often leading to polarization in innovation activities. Large firms tend to focus on exploratory innovation, while small and medium-sized enterprises (SMEs) primarily engage in exploitative innovation (Balsmeier et al., 2024).

Beyond direct corporate taxation, payout taxes—dividend tax and capital gains tax (CGT)—are increasingly recognized for their influence on firms’ strategic decisions. Payout taxes directly affect shareholder returns, which in turn shape firms’ access to external capital. While the literature has extensively studied dividend taxation (e.g., Chetty & Saez, 2005; Yagan, 2015; Becker et al., 2013), CGT remains underexplored, despite its growing relevance as firms increasingly prefer share buybacks over dividends. Recent research by Moon (2022) underscores the importance of CGT, demonstrating that reductions in CGT stimulate corporate real investment. Building on this foundation, our study investigates how CGT reductions influence firms’ innovation activities and strategic directions, particularly by enhancing financial flexibility and enabling greater investment in R&D and high-quality innovations.

To empirically assess the impact of CGT reductions on innovation, we leverage a 2014 policy change in South Korea, where firm classification criteria for CGT rates were redefined. Before 2014, classification was based on sales, employees, capital, and assets, but the reclassification simplified the criteria to sales only, with industry-specific thresholds. Firms previously classified as medium-sized enterprises but reclassified as small and medium-sized enterprises (SMEs) under the new criteria benefited from lower CGT rates (treated group), while firms remaining classified as medium-sized did not experience tax adjustments (control group). Our analysis uses panel data from 2010 to 2019, comprising 2,020 unique firms and 19,525 observations. The treated group includes 227 firms (2,182 observations), while the control group consists of 1,793 firms (14,343 observations). Employing a difference-in-differences approach, we analyze how CGT reductions affected the quantity and quality of firms’ innovation activities.

The results demonstrate that CGT reductions significantly enhance both the quantity and quality of innovation. Firms subject to CGT reductions filed 7.1% more patents than those in the control group, relative to pre-reform levels. Moreover, innovation quality, measured by forward citations, improved by 3.5%, with the most notable increases observed in breakthrough innovations with significant technical impact. Interestingly, while the overall number of unsuccessful inventions remained unchanged, the increase in innovation activity was concentrated in high-impact innovations, suggesting that CGT reductions particularly stimulated quality-driven innovation.

To further explore the mechanisms underlying these effects, we conducted additional analyses. First, we examined changes in firms’ R&D expenditures to determine whether enhanced financial flexibility led to increased investment in innovation. The results confirm that CGT reductions facilitated increased R&D spending, indicating that firms expanded their innovation budgets rather than reallocating existing resources. Second, we analyzed the differential impact of CGT reductions on listed versus unlisted firms. Listed firms, which have greater access to external capital through stock and bond markets, accounted for most of the observed increase in patent activity. This suggests that financial accessibility is a key channel through which CGT reductions influence innovation. Third, we investigated the role of firms’ cash holdings. The findings reveal that cash-rich firms experienced a more pronounced increase in innovation outputs following the CGT reduction, while cash-constrained firms did not. These results align with Moon (2022), who argued that cash-constrained firms prioritize tangible asset investments over R&D when acquiring external capital. Our findings highlight the importance of internal financial resources in enabling firms to leverage CGT reductions for innovation. Finally, we examined the balance between exploration and exploitation in firms’ innovation strategies. The results indicate that CGT reductions encouraged ambidextrous innovation, with firms pursuing both the exploration of new technologies and the exploitation of existing ones. This balanced approach likely contributed to the observed improvements in both the quantity and quality of innovation outcomes (Andriopoulos & Lewis, 2009; Lin et al., 2013).

This study makes three distinct contributions to the literature. First, it extends research on CGT by documenting its novel role in shaping corporate innovation strategies. While prior studies have primarily focused on CGT’s impact on real corporate investment (Moon, 2022), our findings demonstrate that CGT reductions significantly influence firms’ innovation activities, particularly by fostering high-impact, quality-driven innovation. Second, this study contributes to the literature on financial constraints and innovation by providing new evidence that tax policies can enhance financial flexibility and, consequently, innovation outcomes. Building on the work of Brown and Petersen (2011), who emphasized the role of cash holdings in stabilizing R&D investments, we show that CGT reductions particularly benefit cash-rich firms, enabling them to undertake ambitious and costly innovation projects. Third, we bridge two independent streams of research: payout taxes and corporate innovation. Traditionally, the finance literature has examined shareholder return policies, while innovation studies have focused on the determinants of corporate innovation. By integrating these perspectives, we identify a novel channel through which shareholder taxation influences firms’ innovation outcomes. This integrative approach provides a comprehensive theoretical framework linking shareholder return policies, financial flexibility, and corporate innovation performance.

In summary, our findings underscore the critical role of CGT reductions in enhancing firms’ innovation activities, particularly by promoting financial flexibility, enabling ambidextrous innovation strategies, and fostering high-quality, impactful innovations. These insights have important implications for policymakers seeking to design tax policies that support corporate innovation and long-term economic growth.

ABSTRACT. In recent years, tax incentives, mainly provided to firms in the form of tax credits, have become increasingly popular among OECD countries. So that in 2023, out of 38 OECD countries, 33 countries offer R&D tax incentives for their firms, with 21 of them having R&D tax credits. R&D tax credits, reduces the risk of R&D expenses and help stimulate innovation by directly deducting from the firm’s tax liability. Over the past few years, more countries have used this tool to boost innovation within firms. The success of these programs depends on the effective design of their features. Although granting tax incentives has a financial burden (tax revenue loss) for governments, if properly designed and implemented, its fiscal impact is less than other forms of support. Determining eligible activities to receive tax credit, target groups, tax credit rate and permanent or temporary measures, are important decisions that have great effects on program's effectiveness and efficiency. Poor design of these features can lead to resource waste, unintended consequences like rent-seeking, and even market failure. Despite the significance of the issue, there is not enough theoretical and empirical guidance in this field. This research, with a comparative method, has examined the differences and similarities in design and implementation of R&D tax credit programs in developed and developing countries in order to provide suitable policy recommendation for developing countries. The comparative data analysis is conducted in two stages, descriptive and comparative, and success and failure factors were identified in order to extract key policy lessons for improving research and development tax credit systems. The design feature of R&D tax credit policy framework has been extracted from policy reports and global literature, leading to identification of seven key features, each contributing to enhancing effectiveness and efficiency of these programs in promoting research and innovative activities. The feature examined include definition of eligible operations for tax deduction, eligible R&D costs (current, labor, and capital expenditures), basis of credit calculation (volume-based, incremental or hybrid) tax credit), thresholds and caps on R&D expenditures and credit, target group (firm size, firm age, sector or industry, and region), permanent or temporary measures, and specific incentives for small and medium-sized firms. To provide a comprehensive and targeted analysis of R&D tax credits based on these key feature, comparative studies have been conducted, and components have been separately analyzed to clarify various aspects of design and implementation of this policy tool. The proposed policy framework for developing countries in designing R&D tax credit system can be more effective by focusing on four key components. First, focus on firms size as target group, noting the differences between large firms and SMEs. Results of using tax credit are different for two groups, and giving tax incentives can be more cost-effective considering the firm size. SMEs will benefit more from such incentives due to financial constraints and higher risks in R&D projects, and due to knowledge spillovers, they have greater economic effectiveness compared to large firms. Therefore, focusing on supporting SMEs can significantly contribute to sustainable economic growth and increased innovation. Second, there are three approaches to allocating R&D tax credits: volume-based, incremental, and hybrid. For SMEs, the volume-based approach is suitable due to its simplicity, greater generosity, and provides easier access. But for large firms that have more financial resources, the combination of incremental and hybrid approaches can be a suitable option, because this approach helps to better manage operational costs. Third, incentive generosity should be structured in a way that encourages investment in R&D but does not impose too much of a financial burden on government. Developed countries have maintained a balance between effectiveness and efficiency in their policies by setting threshold or cap for R&D expenditures and credits. This strategy also helps developing countries to optimally use public resources to stimulate R&D activities and thus create a sustainable path for economic growth and technological development. Fourth, In the proposed policy framework for developing countries, focusing on improving existing products and services and developing indigenous technical knowledge is more important. Unlike developed countries that are more concerned with innovation at the edge of knowledge, development of new technologies, and global competitiveness, developing countries can benefit from focusing on the operational and practical aspects of R&D to strengthen their indigenous infrastructure and existing capabilities. Thus, rather than pushing the boundaries of knowledge and facing the high risks associated with it, emphasizing on existing technologies improvement and production of new products and processes at the domestic industry level is presented as an effective strategy for achieving sustainable economic growth. In this policy framework, it is preferable to adjust definition of R&D activities to prioritize the improvement of existing technologies and the commercialization of enhanced technical knowledge at the national level. This approach not only controls R&D costs but also increases productivity and effectiveness at operational levels, contributing to sustainable economic development and enhancing regional competitiveness. Additionally, in proposed policy framework for developing countries, focusing on high-tech industries due to their high absorptive capacity and spillover capabilities, along with credit allocation based on firms’ performance and carryforward of unused credits to future years, enhances the effectiveness of tax incentives. This study is ongoing, and we have currently presented part of the findings. We continue to analyze the topics related to policy framework for developing countries, and we hope that by completing this research, we will obtain more comprehensive results that can contribute to improving R&D policies in these countries.

ABSTRACT. Abstract:

A cursory analysis of the electricity access and sustainability transitions data in African countries reveals that most will miss their universal clean energy objectives, set for 2030 under the UN’s Sustainable Development Goals (SDGs), by a wide margin. Meanwhile, the rapid explosion of mobile telephony in Africa since the new millennium has facilitated the emergence of financial technology solutions that have enhanced socio-economic inclusion. Yet, the long-predicted replication of successful mobile penetration in decentralized energy solutions remains elusive. This divergence is compounded by the substantial infrastructure financing requirements for both energy and digital infrastructure. Our objective in this paper is to investigate the differential role that financing domestic technological innovation (endogenous innovative financing), rather than the perennial emphasis on financing infrastructure alone, could play in accelerating the rates of both energy inclusion and the transition from non-renewable to renewable sources. The relevant targets not only relate to those agreed upon by UN member states, but their attainment can also be fruitfully informed by the experiences of selected East Asian states, namely, China and South Korea. Theoretically, we employ the lenses of economic catch-up and mission-oriented financing in order to underscore the urgency of achieving given societal outcomes – “inclusive catch-up”. Methodologically, we employ a trend analysis, apply an energy innovation indicator framework to two countries, and conduct an econometric analysis on innovation financing and energy data. We conclude with policy recommendations for the effective financing of energy innovation in Africa that could yield equitable outcomes much more rapidly.

Introduction:

The centrality of endogenous innovation in tackling the twin challenges of economic growth and the climate crisis is now recognized across mainstream economics, evolutionary economics, innovation systems and sustainability transition theories alike. It therefore follows that the financing of endogenous (or domestic) innovation should enhance the achievement of this dual outcome. However, we also seek to examine the relationship between an intentional investment in renewable energy (RE) and energy efficiency (EE) innovation (as distinct from traditional energy infrastructure financing), on the one hand, and explicitly related societal outcomes, namely, energy access and the transition to cleaner sources, on the other hand.

While the dominant insights from the economics of innovation privilege R&D as a source of (original) innovation, Lee and Walsh (2016) combine those with organizational learning and the sociology of work to demonstrate that even in technologically advanced jurisdictions like the US, the EU and Japan, a large proportion of innovating firms report a significant share of non-R&D innovation (including those that perform R&D). This is of particular importance given that the overwhelming majority of United Nations (UN) member states, which have committed to achieve its Sustainable Development Goals (SDGs) by 2030, are not R&D driven economies. Thus, the notions of secondary innovation (manufacturing based) and tertiary innovation (service driven) take on added significance. Theoretically, we intersect our innovation categories with mission-oriented financing mechanisms for both traditional energy infrastructure (non-renewable) and innovative infrastructure (renewable) to identify the windows of opportunity for inclusive catch-up in RE.

Two SDGs are of explicit relevance to the interrelated objectives of the present study (UN, 2024). SDG 7 calls for sustainable universal energy access, while SDG 9 addresses resilient, sustainable and equitable access to infrastructure, and a significant improvement in innovation capabilities (captured by the number of R&D workers, as well as public and private R&D expenditure). Given that universal electrification has already been attained by some Global South countries, namely, China and South Korea in East Asia, a modified notion of “catching up” with such countries is relevant here. We term this “inclusive catch-up”, which can be defined as “closing the gap in desired human development outcomes, relative to an identified reference society, a specified target, or a given date and at a faster rate than had previously been achieved” (Soumonni & Muchie, 2023, pp. 396-397).

Methods:

1. We provide a trend analysis of energy access data since the year 2000 for Africa, selected East Asian countries and the world. 2. We apply an energy innovation measurement framework developed by Hu et al. (2018) to compare innovation inputs, outputs and outcomes between South Korea and South Africa, which are comparable in population. 3. We specify the data and functional econometric model for testing key variables derived from our conceptual framework. Robustness is verified by comparing pooled OLS, 2SLS and GMM regressions.

Data:

1. Global energy R&D, energy access and RE data from the International Energy Association (IEA). 2. Energy innovation data from South Korean and South African government agencies. 3. Venture capital and private equity investments in RE, EE and clean tech from 49 countries (2014-2022) - African Venture Capital Association (AVCA) data. 4. African Development Bank (AfDB) and World Bank energy sector operations (2014-2022).

Findings and Policy Implications:

1. For Africa as a whole to achieve universal energy access by 2030, it would have to reduce the current gap with the world’s average (90%) or that of China (99% since 2000) much faster than the currently incremental reduction, which is reflective of “policy stasis” (Soumonni & Ojah, 2022). Unless the African continent dramatically increases its current electrification rate (58%) (IEA, 2023b), universal access may not be reached until about 2064. 2. The comparison of energy innovation measures shows that the roughly 10-fold disparity in innovation inputs in favor of South Korea relative to South Africa (e.g. public energy R&D expenditure, RE manufacturing investment, and O&M and services personnel) reflects a commensurate disparity in RE innovation outcomes across the board. 3. The financing of innovation is significantly associated with both energy access and RE generation. Secondary innovation (proxied by manufacturing value added) is also significantly associated with energy access and RE, but not non-RE. 4. Financing mechanisms for RE innovation should judiciously pair instruments (venture capital, private equity, project finance) and actors (DFIs, national governments, private/impact investors, cooperatives) to yield inclusive outcomes. 5. Harmonizing policies through a mixture of bottom-up and top-down initiatives and the financing of Schumpeterian creative destruction across the value chain (seen in East Asia) is insightful for Africa.

ABSTRACT. This project is funded by the NSF (https://nsf.elsevierpure.com/en/projects/eager-pbi-paving-the-path-to-market-the-role-of-place-based-innov) and is the result of a public−private collaboration between researchers at Elsevier-Science-Metrix (David Campbell, PI) and the University of Iowa (Jiajie Xu, CoPI)

Background: Open innovation leverages collaboration to fuel the transition of research and development (R&D) outputs to commercial markets (1). Open innovation is most fully realized in place-based innovation (PBI) ecosystems in which diverse stakeholder types leverage geographic proximity to engage in joint R&D. Geography is a crucial factor in these ecosystems, as it allows for easier connections and the ability to leverage the unique attributes and potential of a specific location to enhance its competitiveness and resilience (2).

Leveraging PBI, the National Science Foundation's Regional Innovation Engines (NSF Engines) program aims to foster and accelerate economic growth to address the nation’s widening regional divide in technological advancements (3). However, there is a gap in research on, and comprehensive data sources for (4), understanding the key drivers and barriers to translating R&D outputs into new products, a key outcome sought through NSF Engine awards. Addressing this knowledge gap could support the adaptive management and success of each NSF Engine.

Goal: By linking large, regionalized data sources covering the entire innovation pipeline from R&D investments to their actualized products in the market, our study attempts—for the first time—an at-scale characterization of both the translation rate of R&D outputs into new products, and the time lag from R&D to commercialization. The study also tests a broad range of potential factors, including those that have been and continue to be subjected to PBI interventions, to address the following questions: • Which of these factors influence the translation of R&D outputs into new products and how? • Which of these factors influence the time from R&D to market and how?

These questions are addressed for each of the key technology and challenge areas (hereafter key technology areas, KTAs ) outlined in the CHIPS and Science Act (https://new.nsf.gov/chips#key-technology-areas-4f3).

Methods: Our study is based on a large-scale data set comprising all US R&D outputs produced since 1996, which consists of close to 6.6 million US publications (mainly peer-reviewed journal articles, conference papers and reviews extracted from Scopus) in the Natural and Applied Sciences, and roughly 1.2 million USPTO patents originating from the US in KTAs (extracted from PatentsView and LexisNexis).

By linking publications to patents, and then the patents to their actualized products in the market (extracted from IPRoduct), descriptive statistics are produced on the translation rate of R&D outputs into new products, as well as on the time lag from R&D to commercialization. The data covers all KTAs, overall and by area, for the US and for NSF-supported R&D outputs (close to 900K NSF-supported publications and 10K NSF-supported patents were identified). Data is disaggregated by US state and county for both groups. It is also sliced according to factors, with a focus on those of highest relevance to PBI, that may influence the commercialization of R&D outputs.

Such factors include, as an example, team size, sectoral and international collaboration, interdisciplinary knowledge and collaboration, gender diversity, entrepreneurship, availability and mobility of workforce, patent quality, local capabilities, availability of financial capital, as well as NSF support codified as PBI relevant or not. Data on these factors originates from a range of sources: Scopus, PatentsView, LexisNexis, IPRoduct, NSF Awards database, NSF-PAR, CHORUS, OpenAire, USAspending, VentureXpert, Business Formation Statistics (Census Bureau), and NamSor. Using this data, it is, for example, possible to compare the share of NSF public–private co-publications linked to patent-protected products to that of all other NSF publications.

The resulting data set is then used to model the influence of selected factors on the targeted PBI outcomes: number of patent-protected products linked to a publication or a patent, and the time lag to product release, either from the date of the related publication or patent. We decided to analyze factors relevant to PBI, rather than assessing PBI interventions, because NSF Engines are too recent and have limited data availability, and other PBI interventions are heterogenous and implemented on a smaller scale, restricting the ability to draw conclusions and/or limiting external validity of the results across the US.

While several of the factors considered in this study can be directly measured at the level of individual publications and patents, which can be regionalized based on the addresses of authors and inventors/assignees, some factors are only available on a regional basis. Two approaches are therefore used to address the study questions: • Multivariate multilevel regression models with some factors measured at the level of individual publications or patents combined with other factors measured for regions (US counties) within which the publications or patents are grouped; and • Multivariate panel regression models with all factors aggregated at the regional level (US counties).

Preliminary results: This study is still in its data collection phase but a complete draft of the descriptive analysis and statistical modeling will be ready for presentation at ATLC25. Our team has already published data on the distribution of patents across KTAs at the level of US states (https://ncses.nsf.gov/pubs/nsb20241/assets/data/swbinv-3.xlsx) and successfully regionalized all patent data at the level of US counties (see Figure INV-4, https://ncses.nsf.gov/pubs/nsb20224/invention-indicators-protecting-useful-ideas#). Additionally, preliminary descriptive results from a test sample for four US universities (University of Washington, Cornell University, Arizona State University and Howard University) show promising results. For all four universities combined, 3,667 patents were linked to products of which 2,108 had a product release date (representing 1.9% and 1.1% of all patents, respectively). Interestingly, the share of patents linked to products was 32% higher for NSF-supported patents (2.5% vs. 1.9%) providing early indication that federal support likely is instrumental in generating R&D outputs with higher market potential and/or in promoting the conversion of R&D outputs into economically valuable products.

References: (1) https://doi.org/10.1142/S1363919622400254 (2) https://www.thecgo.org/research/university-research-technological-commercialization-and-location/ (3) https://doi.org/10.3386/W31541 (4) Manuscript in preparation and shared with our team: https://www.gate.cnrs.fr/évènement/gaetan-de-rassenfosse-ecole-polytechnique-federale-de-lausanne-the-commercialization-of-dod-sbir-patents-a-counterfactual-analysis/

ABSTRACT. Place-based policies are widely employed globally to foster regional economic growth, yet the economic development implications of abruptly terminating these initiatives remain poorly understood. Leveraging a large-scale reform in China as a quasi-natural experiment, this paper examines how the sudden dismantling of Economic Development Zones (EDZs)—triggered by the central government's concerns over waste, fraud, and inefficiency—affected local economic competitiveness and welfare. Urban districts exposed to EDZ closures experienced sustained declines in export competitiveness and worker wages, significantly eroding the economic development benefits accrued during the zones' operational period. Supporting analyses suggest these adverse outcomes primarily arise from disrupted agglomeration economies and increased policy uncertainty resulting from the abrupt withdrawal of government support. However, districts with more innovation capacity, integration into regional supply chains, and trade specialization demonstrated greater economic resilience. These findings highlight the substantial risks associated with abrupt reversals in place-based policies, underscoring the critical importance in the design, placement, and management of zones while operational to sustain regional competitiveness and economic welfare even after explicit policy support is withdrawn.

ABSTRACT. Mechanisms of Firm Learning in the Global South: Political Economy of Quality Infrastructure

Rick Doner Gerry McDermott Dept. of Political Science Moore School of Business Emory University University of South Carolina

Abstract Submitted to the 2025 Atlanta Conference on Science and Innovation Policy (Topic / Area: Forms of governance (e.g., regulation, policy, scientific organization)

Through an analysis of quality infrastructure (QI), this paper addresses the institutional and political challenges of increasing productivity in Low- and Middle-Income Countries. QI refers not to physical infrastructure such as roads and bridges, or even IT-related supports, but rather to a particular form of governance -- the institutional ecosystem that helps firms identify and meet high-quality process and product standards. Concretely, QI institutions, include those, whether national, sectoral or multi-sectoral, or government, private or public-private -- responsible for metrology, standards, testing and quality certification (MSTQ). This list is sometimes supplemented to include accreditation, inspection, technical regulations, personnel certification (and thus training).

Appreciation of these functions has grown in response to concern with disappointing productivity in low- and middle-income country firms. Weak productivity is reflected in the “great divergence,” (unanticipated by traditional development theory), in which poor countries have generally failed to catch up as rich countries continue to pull ahead (Pritchett 1997); it is central to the “middle-income trap” problem (Doner and Schneider 2016); and it is relevant to an understanding of the “missing middle” (i.e. firms operating between a very few large, often multinational, firms and a very large group of small, unproductive firms) and related problem of inequality. The obvious solution to this problem is better adoption of new technologies (Comin and Ferrer 2013). An understanding of QI can help to address two gaps in the literature on technological catch-up by LMIC firms.

First is the nature of the technology relevant to the productivity of these firms. Recent scholarship has acknowledged that relevant technologies are typically not those at the world frontier and more those “new” to the firms, a point captured by the concept of firms’ “capabilities escalator” (Cirera and Maloney 2017; Lall 1992). We thus explore the ways in which QI facilitates incremental learning -- attention and access to diverse sources of applied, experiential knowledge in order to transform existing capabilities to meet basic international quality standards.

Second is the actual mechanisms through which firms might identify and make use of technologies “new” to them. Two traditional answers have proven unsatisfactory. Investments in R&D, while necessary, have been shown to be insufficient. And positive spillovers from MNCs and global value chains have been uneven: helpful to some countries but disappointing in many others where firms “lack the complementary capabilities that would allow them to accumulate knowledge” ( Cirera and Malone 2017: xxii). A key problem involves the ease of knowledge transfer even with relatively standardized systems. Much of the work on emerging markets believes that because MNCs tend to bring more mature, modularized production systems, with discrete packages of technologies and interfaces, then the relevant knowledge and practices are highly standardized and can be “bought off the shelf” or easily replicated by suppliers. (Gereffi, Humphrey & Sturgeon, 2005) However, the work on even apparently highly modularized manufacturing sectors within advanced countries such as automotives, aircraft, and energy equipment has shown increasingly how the diffusion of capabilities for lean production, TQM, and continuous process improvements depends on the tacit knowledge of translating the codified practices from one context to another (Camuffo & Cabigiosu, 2012; Sako, 2004). Moreover, recent research notes how even within MNCs, subsidiaries often have difficulties in such “plug and play” scenarios because they lack the original experiential knowledge of implementation to overcome the ambiguities of cause and effect when developed processes are transferred from one context to another. (Szulanski et al., 2004).

In response to these disappointing spillovers from FDI and GVCs, the mainstream multilateral organizations have increased attention to QI. By incorporating a focus on standards, testing and quality certification into National Innovation Systems and Triple Helix frameworks, the QI literature focus draws attention to the broader institutional ecologies in which local firms identify and adopt “new” technologies.

But if the existing QI literature constitutes important progress on the role of institutions in economic development, it is largely technocratic and supply-driven. As a result, existing analyses say little if anything about puzzles of variation: One involves differences in institutional design: Do all QI systems look the same? Another is the puzzle of performance variation: Why are some QI systems more effective than others? Here we move into more explicitly political concerns: In what ways do broader systems of governance affect performance? What incentives – pressures and opportunities – influence political leaders’ commitment to the challenge of building QI agencies? By addressing these puzzles, the paper aims to fill important theoretical and substantive gaps in the QI literature, while also helping policy makers to identify important opportunities and constraints.

Empirically, the paper draws on a set of comparative cases. Using a combination of most-similar and most-different designs, the paper two sectors (autoparts and wine) in one country (Argentina), and one sector (rubber) in two countries (Thailand and Malaysia).

ABSTRACT. We have previously conducted a series of research profiles (basic and clinical) on COVID-19 (Porter et al., 2020) and a follow-on study of the pattern of Long COVID (LC) research (Porter et al., 2023). Both relied on the standard search protocols provided by the National Library of Medicine (NLM) to identify relevant PubMed papers. Recent research on LC prompts us to strive to enhance that search and apply a new analytical approach to help understand the challenging LC research domain and post-COVID-19 mechanisms.

The NLM LC definition is not fully accepted in the US; in contrast, the UK early settled on LC operational definitions. Fineberg et al. (2024; Figure 1) convey the messiness of specifying what to include as LC-related. They identify 11+ symptoms and 6+ physiologcal systems affected – including cardiovascular, cognitive, kidney, lung, dysautonomia, and autoimmune. We aim to bootstrap from the UK definitions to investigate global research set characteristics. The Long COVID research domain is challenging to specify and to bound. While MeSH terms lag significantly in Pubmed assignment, there are currently nearly 300 articles tagged with the MeSH term “Biomarker.” To illustrate, three biomarkers of note are elevated levels of cytokines such as Interleukin 6; biochemical markers like D-dimer indicating clotting abnormalities; and neurological markers such as elevated neurofilament light chain. We will devise Boolean search strings, with additional delimiters, to identify LC-indicative markers. We then intend to generate a learning model based on these, and query against the entire PubMed corpus. This should generate unique “non-LC” articles that do address LC-relevant research warranting inclusion.

In terms of methodology, this research advances “intelligent bibliometrics” by combining Machine Learning-enhanced analytical capabilities with established policy analytics. The paper then analyzes the expanded LC research data set, seeking to identify research opportunities (gaps) and policy opportunities. We aim to identify potential initiatives to connect topical research across research players -- countries, institutions, and researchers –that are not presently well-linked.

LC research both seeks biomarkers and draws on clinical solutions. However, those are hard to specify for a disease lacking in clearcut biomarkers, distinctive clinical features, or identifiable temporal patterns. Because LC symptoms are diverse and there is no lab test or simple diagnostic approach for isolating LC, the research space will most likely continue to operate inefficiently. Our approach should identify a more inclusive research literature.

Elements of interest in this study include use of Machine Learning trained routines to cluster text terms and phrases, so as to identify vital LC topics. Porter et al., 2023 (Table 7) show topical clusters from the 5,000 LC records compiled through 2022 in the PubMed and Web of Science databases. At the time, 13 topic clusters showed, including, neuropsychiatric, neurological sequelae, fatigue, cognitive deficits, and long-term effects. A simple search of PubMed, as of late 2024, finds over 30,000 papers. We plan to expand that search early in 2025 to more fully cover research publications through 2024. We will reach beyond the bounds of Boolean search on LC by applying the LC knowledge models to be built. The topic clusters will be different and richer than for the earlier results. Denoting changes over time should aid in projecting emerging topics.

Construction of knowledge models entails a several step process – clustering the core of the LC research space using established statistical methods; training a Machine Learning routine in our text analysis software, VantagePoint to form topical clusters; then engaging biomedical experts to check those results. This process aims to produce knowledge models for the major research domains within LC.

These knowledge models would be used to formulate queries within a specially encoded version of the full PubMed database to expand our LC dataspace. The results would be two-fold: expanded search results for knowledge spaces that are difficult to isolate with traditional techniques, plus a collection of topic models that can be used to understand and characterize the LC research space.

Fuller and insightful characterization of LC research is an aim of this study. Because LC symptoms are diverse and there is no definitive lab test or simple diagnostic approach for isolating LC, the research space will most likely continue to be characterized by “lots of loose ends.” Our approach should suggest productive pathways toward integrating this research to enable LC prevention and treatment. By providing a “map” of LC research and by tracking it over time, one could improve research connectedness to advance efforts to mitigate LC. Given that CDC (Centers for Disease Control and Prevention) researchers estimate that, as of August 2024, 17.9% of adults in the United States have experienced Long Covid, LC issues are not going away (Ducharme, 2024), progress is critical.

References

Ducharme, J. (2024), What’s the risk of getting Long COVID in 2024?, Time; [time.com/6999274/long-covid-risk-2024]

Fineberg, H.V., Brown, L., Worku, T., and Goldowitz, L. (Editors) (2024), Committee on Examining the Working Definition for Long COVID, A Long COVID Definition: A Chronic, Systemic Disease State with Profound Consequences, The National Academies Press.

Porter, A.L., Markley, M., and Newman, N.C. (2023), The Long COVID Research Literature, Frontiers in Research Metrics & Analytics.

Porter, A.L., Zhang, Y., Huang, Y., and Wu, M. (2020), Tracking and mining the COVID-19 research literature, Frontiers in Research Metrics and Analytics 5: 594060, 1-18.

ABSTRACT. Background As AI technologies continue to shape critical sectors, the risks associated with AI deployment—ranging from bias and discrimination to unintended system failures—are becoming more apparent (Acemoglu, 2021; Weidinger et al., 2022). Countries are taking different approaches to address AI potential harms through policies and regulations (Elgesem, 2023; Lupo, 2023; OECD.AI, 2021). One area where regulations and policies are being developed is incident reporting (Lupo, 2023), which is seen as a potential tool to prevent repeated AI failures (McGregor, 2021).

Prior work has analyzed incidents database to understand the ethical issues covered (McGregor, 2021; Nasim et al., 2022; Stanley & Dorton, 2023; Wei & Zhou, 2022) and how they could influence policy (Lupo, 2023). However, analysis is still in its infancy, and incident databases have failed to inform AI policies for different reasons such as underdeveloped reporting requirements, lack of standardized taxonomies, etc. (Lupo, 2023; McGregor, 2021; Stanley & Dorton, 2023). Without agreed categories and thresholds, regulators struggle to incorporate AI incident reporting into actual regulation, and companies lack incentives to report and use existing incident databases.

One particular challenge in using incident databases is processing a growing amount of textual information. This is a similar challenge that policymakers face when implementing participatory processes and analyzing civic participatory data. Policymakers are often overwhelmed by large amounts of textual data, reaching the point of cognitive overload (Chen & Aitamurto, 2019). In this domain, it has been explored how to use NLP tools to support policymakers in making sense of civic data (e.g., Arana-Catania et al., 2021; Guridi, Pertuze, et al., 2024; Romberg & Escher, 2022). However, many challenges remain unsolved, ranging from organizational complexities (Guridi, Cheyre, et al., 2024) to technical issues, including the lack of friendly interfaces (Romberg & Escher, 2023). Recent advancements in Large Language Models with chat-like interfaces, such as ChatGPT and NotebookLM, could address existing challenges and support policymakers in analyzing large amounts of text data.

We explore how policymakers could leverage existing LLMs to analyze incident databases and inform AI Policy. Using AI in the public sector could support policymakers’ analysis, but their use has significant risks (Dwivedi et al., 2023; Salah et al., 2023; Weidinger et al., 2022). Since most prior work using NLP analyses summarizing and organizing information, we focus on how LLMs can help policymakers large text databases to existing policy documents. Moreover, we use existing applications that do not require coding, which are more likely to be quickly adopted by policymakers.

Methods Our research question is: How can existing LLMs interphase support policymakers in relating AI incident databases to existing regulations to inform policymaking? To answer this, we conduct a case study using NotebookLM to analyze how incidents in the Artificial Intelligence Incidents Database (AIID) and the AIAAIC Repository relate to the European Union Artificial Intelligence Act (EU AI Act).

We analyze 818 incidents in the AIID, which have the following fields: title, description, date, alleged deployer, alleged developed, alleged harms, and a classification. We also analyze the AIAAIC repository, which has the following fields: title, type, date, country(ies), sector(s), deployer(s), developer(s), system name(s), technology(ies), purpose(s), media trigger(s), issue(s), and a URL to the full web description. For the AIID, we use the title and descriptions, while for the AIAAIC, we use the link to the description.

We use NotebookLM to analyze its ability to analyze incident report databases and relate them to existing policies. We chose this system because it can be trained with specific documents (e.g., PDFs, Google Docs, Websites, and Google Slides) and then be prompted to work only on the uploaded data. The system performed better than ChatGPT in our pilot testing for the specific tasks we aimed to do.

We used a series of prompts so that the system could relate incidents to specific sections in the AI Act and classify them on different criteria such as severity, likelihood of violation, etc. We manually revised if the resulting analysis corresponded with the data in the incidents’ description and the AI Act. We documented the prompts and results alongside notes about the system's performance.

Findings and Discussion Our findings demonstrate that out-of-the-shelf LLMs, such as NotebookLM, can assist in classifying AI incidents under the framework of the EU AI Act by identifying links between reported incidents and relevant provisions. For example, incidents involving deepfake scams and the use of AI for generating illegal content align clearly with Prohibited Practices under Article 5 of the EU AI Act. Similarly, using AI-generated voice clones without consent raises concerns related to Transparency Obligations outlined in Article 52, while anti-shoplifting systems potentially fall under High-Risk Systems due to their use of biometric identification and potential biases, as detailed in Annex III. These findings underscore the potential of LLMs to classify incidents based on risk levels and regulatory relevance, even with limited contextual information.

However, the analysis also reveals critical gaps in current incident reporting practices. More detail in incident descriptions limits their utility for direct application to regulatory frameworks, highlighting the need for standardized reporting criteria and comprehensive taxonomies. Despite these challenges, LLMs show promise in reducing the cognitive burden on policymakers by augmenting their ability to link incidents to regulatory provisions. This capability enhances the efficiency of regulatory analysis and ensures a more systematic approach to evaluating AI risks and their policy implications.

We faced several limitations that must be addressed before the widespread use of LLMs in this kind of analysis. For example, NotebookLM could not analyze tables or spreadsheets, forcing URLs or PDFs to be used, making the analysis less accurate. Moreover, more analysis is needed into how the LLMs use different levels of certainty in their response to inform people and how policymakers could systematically analyze that.

ABSTRACT. The transition to green technology is critical for addressing climate change, yet our understanding of how this transition occurs across different stages of production processes remains limited. While existing literature has extensively analyzed green technology adoption at the industry level (e.g., Acemoglu et al., 2012; Feng et al., 2022) or focused on specific technologies (e.g., Liu et al., 2023), these approaches mask important variations in transition patterns across different production process stages. Traditional industry-level analyses face three key limitations in understanding green technology transitions. First, they overlook how different stages within the same industry face varying barriers to innovation adoption - a critical oversight as technological challenges often vary significantly between basic materials development and final system integration. Second, they fail to identify potential technological synergies across similar processes in different industries, missing opportunities for cross-sector learning and technology transfer. Third, they provide limited guidance for designing targeted policies that address specific technological bottlenecks in the production process, potentially leading to inefficient resource allocation in technology development programs. This study introduces a novel process-level framework for analyzing green technology transitions, focusing particularly on the battery sector as an exemplar case that clearly demonstrates the value of this approach. By examining transitions at the process level - from upstream (raw materials and resources) through midstream (core component manufacturing) to downstream (system integration and application) - we can better understand where technological bottlenecks occur and how they might be addressed. Our analysis leverages advanced natural language processing techniques applied to 3.2 million patents, providing unprecedented granularity in mapping technological evolution across process stages. The importance of this research lies in its potential to inform more effective policy design. Understanding how green technology transitions vary across process stages can help policymakers develop more targeted interventions and optimize resource allocation. The battery sector, representing 28% of our sample and showing clear process-stage differentiation, provides valuable insights for developing stage-specific policy approaches. While our detailed analysis focuses on the battery sector, our methodological framework offers a new lens for examining technological transitions more broadly.

Our analysis framework consists of three distinct process stages: upstream processes involving raw materials and resource conversion, midstream processes focused on core component manufacturing, and downstream processes covering system integration and application. This three-stage framework provides systematic structure while maintaining sufficient granularity to identify specific technological bottlenecks. The framework is particularly well-suited for analyzing the battery sector, where we can clearly observe these stages from material development through cell manufacturing to system integration. We construct our dataset from USPTO patents granted between 1990 and 2023, identifying green technology patents through the intersection of three major classification systems: OECD's ENV-TECH, WIPO's Green Inventory, and EPO's Y02/Y04S tagging schema. This comprehensive approach results in a dataset of 3.2 million patents, with battery technology representing 28% (approximately 896,000 patents). For each patent, we collect and combine title, abstract, and citation information to create a rich text corpus for analysis. Our methodological approach leverages BERT for Patents, a specialized language model pre-trained on USPTO patent documents, which offers several advantages for our analysis. First, it better captures patent-specific terminology and technical context through its training on patent documents. Second, it understands hierarchical relationships between technical concepts, enabling more nuanced classification of process stages. Third, it can identify subtle technological relationships that might be missed by traditional keyword-based approaches. The analysis proceeds in three steps. First, we generate patent embeddings using BERT for Patents, capturing the semantic content of each patent's title and abstract. Second, we apply k-means clustering (k=30, determined through silhouette analysis) to these embeddings, as this method efficiently handles our large-scale dataset while providing clear, interpretable clusters. Finally, we map these clusters onto our three-stage process framework based on their technical characteristics and position in the value chain, validated through manual review of representative patents from each cluster. Initial clustering results reveal distinct patterns in the battery sector. We identify clear clusters for material development (clusters 29 and 8, comprising 1,366 patents), cell manufacturing (clusters 15, 22, and 4, totaling 3,104 patents), and system integration (cluster 17, containing 2,231 patents). The varying sizes and distinct characteristics of these clusters provide a robust foundation for analyzing process-stage differences in green technology innovation.

Our analysis of the battery sector reveals how different stages of the production process face unique technological challenges: Upstream processes (materials development, ~1,400 patents) show highly specialized innovation focused on fundamental research, particularly in metal materials and electrolyte development. These clusters reveal the deep technological changes required at the material level for green transition. Midstream manufacturing (electrode and cell manufacturing, ~3,100 patents) represents our largest cluster group, highlighting the critical challenge of scalability. These patents focus on translating basic materials into manufacturable components, addressing diverse technical challenges in electrode design, conductive structures, and assembly processes. Downstream integration (battery systems, ~2,200 patents) demonstrates a more application-oriented innovation pattern, focusing on practical implementation challenges including battery management systems and pack design. This stage reveals how new green technologies must be effectively integrated into existing systems. The varying cluster sizes and technical focuses across stages suggest fundamentally different innovation dynamics requiring distinct policy approaches. This evidence demonstrates why examining green transitions through a process-stage lens is crucial for effective policy design.

This research makes three principal contributions to understanding green technology transitions. First, methodologically, we demonstrate how combining a process-level framework with advanced NLP techniques can reveal the multi-layered nature of technological transitions. Second, empirically, we provide detailed evidence of how green innovations manifest differently across process stages, from fundamental material changes to system-level adaptations. Third, we show why considering process stages is crucial for designing targeted policies that can effectively address specific technological bottlenecks in the green transition. Our findings suggest that effective green transition policies must be tailored to address the distinct challenges at each process stage - from supporting fundamental research in materials development, to facilitating manufacturing scale-up, to enabling system integration. This nuanced understanding can help policymakers develop more targeted interventions across the production process.

ABSTRACT. Service robots are a key part of emerging technologies and drive innovation across industries. Their rapid growth and interdisciplinary nature make measuring and classifying them difficult. This study develops a framework to measure SR technologies. It focuses on two questions: (1) How can we create a reliable measurement framework for SR technologies? (2) How does combining human judgment with AI, such as ChatGPT, improve the identification and classification process? The study begins with an initial dataset of 174 firms identified using the International Federation of Robotics (IFR) categorization. ChatGPT was asked to generate lists of 50 SR-related firms in multiple iterations, resulting in 429 unique firms. These results were compared to a refined list of 48 firms, finalized through manual expert review. ChatGPT showed scalability but often included duplicates or irrelevant firms. In contrast, human review ensured contextual accuracy, particularly for firms without patents or explicit SR references. To link product descriptions to patents, four embedding models were tested: (1) Word BERT with individual word embeddings, (2) Word BERT with averaged word vectors, (3) Sentence BERT, and (4) Patent BERT. A reliability test on 20 random sample patents showed that Patent BERT produced the best results regarding precision and contextual relevance. Using Patent BERT, IFR product descriptions were matched to relevant patents through cosine similarity. Patents with scores above 0.65 were assigned to IFR categories. Patent data was sourced from PatentsView, which focused on 100,000 patents filed between 2021 and 2023. The resulting treatment group of SR patents was used to train a supervised model. The model classified patents into SR and non-SR categories. Stratified 5-fold cross-validation ensured balanced data and robust performance evaluation. This framework provides a reliable way to measure SR technologies. It addresses gaps in traditional methods by combining human expertise with AI tools like ChatGPT and Patent BERT. The approach ensures accurate and scalable measurement, making it applicable to other emerging technologies.

ABSTRACT. One of the strongest trends in the innovation measurement literature is the critique and consensus that we need new, contextualised, policy relevant indicators (Gault 2008; Iizuka and Hollanders 2017; Teng-Zeng 2009). While there is much experimentation with the underlying theoretical and methodological frameworks (Archibugi and Coco, 2004; Borras and Edquist, 2013; Godin et al., 2021), few alternative perspectives and new approaches emerge from, or focus on the distinctiveness of innovation in developing country economies.

Indeed, the structure of low and middle-income economies differ markedly from the high-income countries on which most of the innovation theory and literature is based, and their insertion into global value chains is constrained. The result is likely to be that the national STI policy mixes (Borras and Edquist 2013; Cocos and Lepori 2020) adopted are largely borrowed from international practice and are not oriented sufficiently to the realities of developing country conditions (see Borras, 2011; Boon and Edler, 2018; Hollanders and Arundel, 2007), nor grounded in an understanding of local firm innovation and production capabilities.

The literature on developing country innovation measurement tends to reflect normative conceptualisations of innovation systems characterised in terms of ‘absence’, of what does not exist, when measured using the standards and models developed in high-income countries (Muchie et al 2003; Pietrobelli, 1998). This is evident for instance, in attempts to explain “the failure to compete” (Lall et al 2002). This paper contributes to open-up and broaden research on new analytical perspectives and frameworks to understand firm innovation in a middle-income country context, and hence, what should be measured, to understand how firms contribute to economic growth.

The paper therefore begins from the assumption that appropriate and relevant innovation indicators need to be informed by context-specific evidence on the determinants and nature of innovation, technological learning and upgrading in firms (Edler 2009; Freeman and Soete 2009; Lall 1993; Lorentzen 2009; Zanello et al 2015). Specifically, it demonstrates a model and methodology that can be used to map the modes of innovation and profile patterns found across local firms, to provide more valuable indicators and evidence to inform policy mixes.

A research stream on modes of innovation was initiated by the OECD in 2009, to develop a taxonomy of firms with distinctive sets of capabilities that require different kinds of policy support. This paper initially draws on experimentation (by Kruss, Kahn and Petersen (2024)) using innovation modes in South Africa by applying the conceptual taxonomy of output indicators developed by Arundel and O’Brien (2009) to an analysis of the South African Business Innovation Survey (2014-2016) dataset. The application of such multi-dimensional frameworks, that considers a wider range of heterogenous firm capabilities, can have great value but has been limited in the developing country context.

To categorise firms, product innovators were grouped according to their market reach, product novelty and responsibility for development, yielding five mutually exclusive groups: ‘new to market international innovators’, ‘new to market domestic innovators’, ‘international modifiers’, ‘domestic modifiers’ and ‘adopters’. We then mapped patterns of characteristics, including sector, use of advanced technologies, firm size and ownership, to profile the firms in each category.

The analysis indicates that the set of South African firms with cutting-edge innovation capabilities (‘new to market international innovators’) is relatively small (13%), and the group of ‘new to market domestic innovators’ (7%) is even smaller. Just over a quarter are classified as ‘adopters’, whose innovations are developed by external organisations (28%), ‘international modifiers’ make up 34% of product innovators, and ‘domestic modifiers’ 19% of product innovators.

This approach has limitations, in its use of a single top-down output oriented taxonomic methodology, and an old dataset. The categorisations were borrowed from the Australian context, and therefore were not necessarily suited to the South African business sector. The analysis was therefore extended, to conduct a data driven analysis on more recent data. The data driven analysis is a bottom-up approach, which entails using exploratory factor analysis to identify more inherent groupings of firms in the data. The resulting dimensions that identify the groups are more reflective of the true structure of the South African business sector and its challenges.

Data from the South African Business Innovation Survey (2019-2021) was used to conduct the factor analysis. The analysis identified a number of groups of firms that could explain the interrelationships among several observed variables. The observed variables included both inputs into innovation as well as innovation outputs. This included types of innovation activities, intellectual property rights, use of advanced or emerging technologies, barriers to innovation, sources of information, collaboration, types of innovations introduced and the outcomes of their innovations. Business characteristics such as size, skilled employment, and part of an enterprise group, were also included.

The results show the value of using more complex indicators of innovation modes, profiled in a more differentiated manner, to critically adapt and design industrial, sectoral and innovation policy mixes grounded in the understanding of national innovation capabilities and the practices of firms. The analysis highlights the need to build technological capabilities across the full range of firms, including those displaying no or minimal innovation capabilities. Defining modes can provide greater nuance to inform policy interventions than is possible from the binary indicators that are typically used. A multi-dimensional taxonomy can equally be used to track change over time, towards desired policy goals. The challenge for the design of a STI policy mix is to shift and orient the patterns of modes of innovation in multiple directions, over time, towards the desired development goals, targeted to and informed by the full range of firm-level capabilities and needs.

ABSTRACT. This paper explores the link between innovation and Global Value Chain (GVC) participation, especially focusing on how Africa could strengthen its GVC integration. Through a systematic search and review of existing research, the study emphasizes the significance of National Innovation Systems (NIS) for technological progress and potential industrial growth in Africa. Despite Africa facing challenges like weak research infrastructure, limited funding, and skills gaps, there are opportunities for leveraging innovation to enhance GVC participation.

The research aims to establish a foundation for Africa to boost its GVC integration through innovation, potentially accelerating industrialization. A “systematic search and review” method identified and evaluated relevant literature, focusing on innovation, GVCs, and industrialization in Africa. The study involved selecting key academic databases, using specific search terms related to innovation and GVCs, and setting criteria to include relevant qualitative and quantitative studies while excluding irrelevant sources.

Data analysis involved a narrative synthesis to pinpoint themes such as the role of government policies, technology adoption, and collaboration within GVCs. By identifying patterns across the literature, the study highlighted how governments and firms must work together to enhance Africa’s GVC involvement. Strategic government interventions like industrial policies and infrastructure investment are essential, along with proactive strategies by African firms, such as upgrading technology and expanding exports. These can make African businesses more competitive globally.

The study also drew lessons from countries like China and South Korea, which successfully integrated into GVCs through government support and innovation promotion. African governments can take similar steps by fostering research and development, technology transfer, and creating innovation hubs to accelerate growth.

At the firm level, strategies such as technology upgrading, strategic alliances, and capacity-building programs are recommended to improve Africa’s GVC participation. Export activities are essential for African firms to access new technologies and global best practices. The paper emphasizes the importance of cross-border collaboration and regional partnerships in fostering innovation, sharing knowledge, and boosting industrialization.

Key research gaps identified include mapping African GVC participation, understanding GVC dynamics in Africa, promoting technology adoption, and evaluating policy frameworks supporting innovation. The study proposes a research agenda focused on fostering innovation-led growth within African GVCs, with priorities such as enhancing technology adoption, improving policy support, and promoting cross-border collaboration. By addressing these areas, Africa can better position itself within global networks, promoting sustainable economic growth.

ABSTRACT. The role of high-tech industrialisation as an important driver of economic growth and resilience, global competitiveness, and modern industrial development is well recognised in the development economics literature. Several technology-intensity classification systems have thus been developed, classifying manufacturing sectors into different groups, largely based on their R&D intensity or expenditure. Given the degree of heterogeneity across sectors, particularly in terms of R&D expenditures, there is a need to examine whether these classification systems are fit for purpose in the context of rapidly changing manufacturing systems, particularly in developing countries. This paper presents a perspective that proposes innovation indicators to measure innovation in technology-intensive manufacturing industries in South Africa accurately. To this end, this study reviewed key existing technological classification systems and examined their importance to the South African economy. Next, all identified technology-intensive classification systems were mapped to South Africa’s industrial structure using sectoral data from the Quantec database and proposed a new set of indicators to better measure and monitor innovations in these industries. Finally, the proposed indicators were operationalised using data from the 2020 WBES for South Africa. The findings reveal a heterogeneity of indicators for capturing innovation in South Africa. Also, the observed heterogeneous performance of HT industries across the selected innovation and socio-economic development indicators highlights the unique complexities in the innovation-technology intensity-development economics nexus. These findings point to the nuanced landscape South African firms operate in, necessitating even a much more equally nuanced policy response. The findings underscore the need for a multifaceted policy approach that recognises the distinct challenges and opportunities in the South African manufacturing industry. It further shows that South Africa’s development strategies need not only focus on fostering innovation in high-tech industries but also leverage the capacity of medium- and medium-low-tech industries for balanced and inclusive economic growth.

ABSTRACT. Beyond its historical role in economic growth, innovation now takes on new dimensions, becoming essential in addressing contemporary issues (OECD & Eurostat, 2018). For innovation policies to contribute to economic growth, social progress, gender equality, and environmental performance, participatory Research and Development (R&D), which strengthens collaboration and engagement among diverse stakeholders, has to take these issues as part of their design, targets, monitoring and assessment. (Woolthuis et al, 2005; Elisa Arond et al., 2010; Galvao et al., 2019). Furthermore, a transformative model of R&D and innovation must take into account the differences between countries in the Global South and North. As the world grapples with social and environmental issues, developing countries often need to deal with added hurdles that developed nations have already overcome (Lundvall et al., 2009; Mohamed et al., 2022). Brazil faces persistent challenges in effectively leveraging private R&D investments. Despite longstanding government efforts, including tax incentives and economic subsidies, these measures have proven insufficient (Pacheco & Corder, 2010; De Negri, 2017; MCTI, 2021) to enhance the country's technological capabilities, drive innovation, and address societal challenges. In 2013, the Brazilian government founded the Brazilian Company of Research and Industrial Innovation (EMBRAPII), a private non-profit organisation designed to promote collaborative research between Research Organisations (ROs) and industry. This unique model within the country, inspired by the French Carnot Institute and the German Fraunhofer (Castro et al., 2017; Salles-Filho et al., 2021), holds significant potential to serve as a framework for addressing and overcoming this context of deficiency. EMBRAPII receives funds from the Federal Treasury to support selected and accredited Research Organizations (ROs). After accreditation, these ROs are considered EMBRAPII Units (EUs). The EUs must pursue partnerships with companies to develop pre-competitive technological innovations aligned with the company's interests. Each project receives one-third of its funding directly from EMBRAPII, with the remaining portion funded by the company itself plus a counterpart from the EUs. On average, companies have contributed 50% of the investment per project. The EMBRAPII model imposes governance schemes to enhance effectiveness. Funds are released only after the EU signs a contract with a company to develop specific projects, which are monitored by both parties and require company approval before completion. Upon project conclusion, an external expert reviews the process. In addition, EMBRAPII also provides training on project and innovation management and rewards performance. We conducted an evaluation of the projects under the EMBRAPII model, with the specific interest of understanding the generation of innovation, the role of company size in this process, and the outcomes of these innovations for the companies in a multidimensional approach, measuring economic, social and environmental outputs, outcomes and impacts. After an extensive literature review, we analyzed exclusive data provided by EMBRAPII, which allowed us to define a sample of 940 projects completed between 2015 and 2023, carried out by 57 EUs and 718 companies, with 50% comprising small and micro-enterprises, and 41% medium and large companies (9% could not be categorized). Adding to that, we administered a questionnaire to the participating companies, yielding responses from 211 of them (29.5% of the sample). Through the descriptive statistical analysis of data, and with the support of input-output analysis, cost-benefit analysis, and Qualitative Comparative Analysis, we seek to assess the impacts of Embrapii using the following indicators: generated innovations, with an emphasis on the percentage of innovations that reduced environmental impact; EMBRAPII's contribution to achieving innovation (redundant causality factor, as proposed by Salles-Filho et al., 2011); return on investment; generation of technological capabilities; and propensity to invest in R&D before and after the project. All indicators were analyzed with a segmentation by company size. Our preliminary results indicate that 72% of projects conducted under the EMBRAPII model generated at least one innovation, of which 36.5% led to a reduction in environmental impacts. Furthermore, 51% of the companies said that the model contributed significantly to achieving innovation; 37% said they would not have undertaken the project without EMBRAPII funding, and 19% said they would have done it anyway. Our study also indicates that for every 200 thousand dollars invested by EMBRAPII, 450 thousand dollars are added to the Brazilian GDP and 33 jobs are created; and for every dollar invested by companies, 2.7 dollars are returned to them. While this research is still ongoing and in-depth analyses are being finalized, we can already affirm that the EMBRAPII model has a positive impact on innovation generation in partner enterprises. Our preliminary data points to a greater impact of Embrapii on small and micro businesses: 85% of them achieved innovations, while this importance drops to 62% for medium and large companies. Medium and large companies also reported more frequently that they would have carried out the project regardless of the funding received. In a country where the industry is traditionally less innovative and small businesses face pronounced barriers, collaborative models that include small and micro enterprises and are implemented under a broad governance framework (encompassing not only financial and technical oversight but also managerial governance) may present a new path of effectiveness in innovation policy.

Main references: Elisa Arond, A. Ely, Martin Bell, M. Leach, Ian Scoones, & Andy Stirling. (2010). Innovation, Sustainability, Development: A New Manifesto. STEPS Centre. Lundvall, B.-Å., Vang, J., & Joseph, K. J. (2009). Innovation System Research and Developing Countries. Em B.-Å. Lundvall, K. J. Joseph, C. Chaminade, & J. Vang (Orgs.), Handbook of Innovation Systems and Developing Countries. Edward Elgar Publishing. Mohamed, M. M. A., Liu, P., & Nie, G. (2022). Causality between Technological Innovation and Economic Growth: Evidence from the Economies of Developing Countries. Sustainability, 14(6), 3586. OECD & Eurostat. (2018). Oslo Manual 2018: Guidelines for Collecting, Reporting and Using Data on Innovation, 4th Edition. OECD. Salles-Filho, S., Bin, A., Bonilla, K., & Colugnati, F. A. B. (2021). Effectiveness by Design: Overcoming Orientation and Transaction Related Barriers in Research-Industry Linkages. Revista de Administração Contemporânea, 25(5), 22. Woolthuis, R. K., Lankhuizen, M., Gilsing, V. A system failure framework for innovation policy design, Technovation, Volume 25, Issue 6, 2005.

ABSTRACT. High-challenge, innovation-driven R&D initiatives—marked by complexity, uncertainty, and significant potential impact—require a customized evaluation framework that goes beyond traditional assessment methods. Despite recent policy adjustments, such as removing penalties for rigorously conducted but unsuccessful projects, researchers still perceive limited support for ambitious, high-risk research within government R&D systems. Standard evaluation frameworks often fail to capture the unique attributes of high-risk R&D projects, resulting in a lack of proper recognition for projects with transformative potential.

To address this, we propose a formative, mission-oriented evaluation framework that aligns with evolving policy and technology landscapes, thus enabling projects to achieve mission-specific objectives and generate high-value outcomes. Our framework builds upon best practices from national fiscal evaluations and international methodologies, introducing assessment criteria that emphasize goal setting, risk management, milestone tracking, and adaptable governance. This tailored approach prioritizes key areas including research planning (goal setting, portfolio strategy, adaptability), execution (collaborative governance, monitoring), and outcome generation (scientific, economic, societal impact, and unanticipated benefits).

To validate this model, we suggest pilot implementations in strategic technology projects, such as next-generation semiconductors or advanced nuclear technologies. This customized evaluation framework aims to foster an environment conducive to breakthrough innovations, ultimately enhancing the alignment and effectiveness of national R&D investments toward strategic objectives.

ABSTRACT. 10th Atlanta Conference on Science and Innovation Policy

Portfolio-based Cost-Benefit Methods for Evaluating Large Research Portfolios

Jonathan Merker RTI International, Chicago

Amanda Walsh RTI International, Chicago

Background: Solow (1957) introduced technological change as being a relevant factor for productivity growth, and thus long-term economic growth. Under this model, research and development (R&D) is a key driver of productivity, which is a topic that has received great attention in economic literature.

Within this literature, role of government in pursuing R&D has been explored. The main rationale for government investing in R&D is government fixing market inefficiencies that lead to sub-optimal private investment in R&D (Arrow, 1962; Nelson, 1959). While there are many justifications for government intervention regarding market inefficiency, some notable justifications are that government has economies of scale in investing in large infrastructures, that government can burden more risk in dealing with technical complexity, and that government can invest in research with long time horizons to achieving commercial benefits (Tassey, 2004).

The Commonwealth Scientific and Industrial Research Organisation (CSIRO) is Australia’s national science agency. Defined by its governing legislation, CSIRO is intended to be an applied science organization, support industry and the objectives of the Australian public, and encourage the utilization of scientific output (Science and Industry Research Act, 1949). CSIRO invests broadly and holistically across the spectrum of innovation to deliver impact, competitiveness, health and security, and environmental stewardship to current and future generations of Australians.

Approach: Across its research portfolio, CSIRO ensures that it is delivering value to Australia through the regular performance of impact case studies using cost-benefit methodologies. CSIRO also commissions bi-annual value reports to assess the collective impact measured by the studies. The bi-annual value report utilizes all existing case studies as data points for CSIRO impact, assessing impact across CSIRO as well as by year and within individual research disciplines. Roughly 25 case studies are completed per year, with selection to ensure that each of CSIRO’s business units is assessed. To date, each case study was completed either by CSIRO’s internal impact evaluation team or by one of six external institutions.

CSIRO’s mission fits well within the conceptual role of government in R&D. CSIRO invests in large research infrastructures, technologically risky projects, and other research portfolios that may deliver impact over long time horizons, which industry is not well suited for. However, case studies with long time horizons provide for difficulties in evaluation. Many case studies are dependent on analysis of benefits that accrue to Australia through the commercialization process, which can be difficult to project for nascent research projects.

RTI International prepared case study analysis for the 2020, 2022, and 2024 Value of CSIRO reports. For each, we reviewed and updated every case study in the portfolio (N=139 in 2024) to homogenize inflation and discounting methods. We generated an aggregate time series of benefits and costs across all studies that monetized the value of activities initiated in the last 25 years (N=83 in 2022). To reduce uncertainty, we removed benefit and cost projections beyond 10 years for each study. We used the resulting time series to estimate Australia’s return on investment in CSIRO’s activities covered in the portfolio.

There are multiple benefits to performing case study aggregation and assessing CSIRO’s value at the portfolio level. As the portfolio of studies grows, each bi-annual value assessment provides a more robust picture of CSIRO’s impact. The consistent application of assessment methods across contractors, verified through rigorous quality monitoring, enables case study comparability. Applying consistent assessment methods to the bi-annual value reports further increases the reliability of aggregate impact estimates.

Beyond these benefits, the portfolio method is uniquely suited to remedying the difficulties of assessing long time horizon case or technologically risky studies. Across technologically risky projects, it is commonplace for a candidate line of research to be deemed infeasible and discarded. However, these failed technologically risky products are rarely the subjects of evaluation. Evaluation at the portfolio level allows these cases to be incorporated into overall return on investment statistics without individual case studies being performed.

Technologically risky projects are often the projects that yield breakthrough innovations that provide large downstream returns. Across CSIRO case studies, it is not uncommon for a given individual study to have a return on investment above $100 per dollar invested. Evaluation at the portfolio level also diminishes the impact and the mischaracterization of such outliers.

Finally, for long time horizon projects, the removal of future benefits beyond 10 years seeks to reduce the impact of uncertainty in the overall estimates of CSIRO impact. While this does often reduce overall return on investment figures, the growing corpus of studies includes retrospective case studies from the same technology areas as current prospective long time horizon studies, diminishing the negative effects of reducing included value-years.

Due to the design of the value report approach for CSIRO and its perceived advantages for research translation, we believe that the portfolio methodology could be used for performance evaluation of other research translation agencies or technology accelerator programs.

References: Arrow, K. J. (1962). Economic Welfare and the Allocation of Resources for Invention. In C. K. Rowley (Ed.), Readings in Industrial Economics (pp. 219–236). Macmillan Education UK. https://doi.org/10.1007/978-1-349-15486-9_13

Nelson, R. R. (1959). The simple economics of basic scientific research. Journal of Political Economy, 67(3), 297–306.

Solow, R. M. (1957). Technical Change and the Aggregate Production Function. The Review of Economics and Statistics, 39(3), 312–320. https://doi.org/10.2307/1926047

Tassey, G. (2004). Underinvestment in Public Good Technologies. The Journal of Technology Transfer, 30(1–2), 89–113. https://doi.org/10.1007/s10961-004-4360-0

ABSTRACT. Research funding agencies have a unique role in the research ecosystem. By providing the financial support necessary for research to be conducted, funders provide the foundation on which the research ecosystem is built and help shape the directions, processes, and policies though which research is conducted. But because most funders distribute funds collected from external sources, funders are consistently asked to demonstrate that they are responsible stewards of the funds they manage and that the research conducted with those funds produces tangible benefits to society. Funders therefore need comprehensive and actionable strategies for assessing the impact of the research they support.

In this proposal, I present an overarching framework proposed for assessing biomedical research impact at the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), one of the 27 Institutes and Centers comprising the US National Institutes of Health (NIH). The framework currently consists of three components: a logic model of NICHD operations, a suite of associated indicators, and a data collection and analysis approach for generating these indicators. Although developed specifically for NICHD, the strategy is adaptable for other biomedical funding agencies.

The first component in the strategy is a logic model depicting the inputs, outputs, short-term outcomes, and long-term outcomes of all project types supported by NICHD. In the model, investigators use knowledge and resources to apply for various types of project funding. Awarded projects then produce outputs like knowledge, people, and products. Those outputs then lead to a series of short-term research outcomes like an informed research community and a diverse and skilled workforce and short-term health outcomes like new or improved treatments and better health behaviors. Finally, these short-term outcomes lead to NICHD’s long-term outcomes of producing new research and improving human health. In this framework, impact is envisioned as a series of connections, or pathways, through the model. Impact can therefore be achieved in multiple ways through multiple channels depending on the type and goals of each project. Impact for a training project will be different than impact for a research project because the goals and intended impact pathways are different, but both pathways ultimately lead to NICHD’s long-term outcomes of new research and improved health.

The second component in the strategy is a suite of quantitative indicators that indicate progress through the model. For example, article-to-article citations indicate that knowledge produced by a project informed the research community and was then used to produce new research. Trainees actively publishing in the academic literature indicate that people joined the research workforce and are now contributing to the production of new research. Article citations in patents and policy documents indicate that scientific discoveries resulted in new technologies, treatments, and/or policies that can improve health. A range of indicators are therefore needed to capture the myriad impact pathways depicted by the model, and different indicators are needed for different impact pathways. These quantitative indicators will be complemented by qualitative inputs like case studies and impact stories to maximize comprehensiveness and minimize bias.

In the strategy, these indicators will be used to assess large funding programs and to identify significant research and health outcomes but will not be used to assess individual projects or investigators. Following Thomas Kuhn, the strategy considers significant research impacts to be rare occurrences that resist aggregation. Aggregate impact measures imply that every research output produced should be impactful – that every publication, trainee, or product should change the world – which is neither realistic nor fair. Most scientific research results in incremental advances that collectively enable substantial, groundbreaking discoveries to be made. From this perspective, specific investigators or specific projects rarely change the world, so they should not be expected to do so. Nor should anyone expect every research output they produce to do so. Instead, the proposed indicators will be used as signals that significant research or health outcomes may have occurred at a specific time and place.

Since NICHD supports over 3000 projects per year, the third component in the strategy, the data collection approach used to generate these indicators, will rely heavily on automated scripts and application programming interfaces (APIs). These processes will typically begin with funding data collected from NIH administrative data systems like RePORTER. A series of APIs will then be used to collect data on outputs like people, publications, patents, etc. resulting from NICHD funding and then collect outcome data like citations in various document types that are associated with those outputs. Finally, analysts will identify findings, projects, or programs with high outcome measure values as candidates for in-depth follow-up. Investigator-supplied progress reports and program staff nominations will be used to supplement and validate the indicator-derived findings. Finally, analysts will collaborate with program staff to write impact narratives for the significant results or programs that describe the context, results, significance, and NICHD involvement for each one.

This proposed strategy is currently under review by NICHD leadership and staff. The strategy proposes substantial philosophical and operational changes to how impact is currently assessed at NICHD, so if it were to be adopted, it would take time for these processes to be implemented and for stakeholders to adjust. But I hope its benefits will be worth the cost. By shifting impact assessment away from individuals and projects, the strategy aims to reduce pressure on NICHD’s funded investigators to demonstrate the productivity and significance of their work, allowing them to focus more on the work itself. And by expanding the definition, and partially automating the process of identifying, impact, it also aims to enable NICHD to identify, communicate, and celebrate more of the life-changing research that it supports. NICHD funding supports the research enterprise and saves lives. I hope the proposed strategy helps it tell that story.

ABSTRACT. Social capital incorporates the social process and status procured from the intricate structures of an individual's social network (Bordeau,1986). Social connections are a resource, and, as with many resource-based advantages, with whom we connect and when can make it easier to procure further advantages. This is true for various individuals (e.g., in business and other organisations), but also very true for academic researchers. Those who undertake an academic career will understand that 'social capital' is a significant precursor to initial and continued academic success (Abassi et al., 2011). However, success can mean many things, for example, acquiring research funds, scientific puzzle solving, and/or advancing a university’s mission (Ioro et al., 2017). At the very least it could be as simple as “becoming familiar with alternative ways of thinking and behaving” and being “at a higher risk of having good ideas” (Burt, 2004, p. 349).

The aim of our research is to examine how scientists’ collaborative preferences and strategies vary by groups, and how they are realized in actual collaborative behaviors. We focus specifically on academic faculty researchers who are embedded in institutions that provide both resources and rule structures important for career outcomes and advancement. How individual scientists prioritize the type of social capital that they seek in collaborative relationships may have an effect on the outcomes and visibility of their work and their careers overall.

This research advances an initial study, whereby collaboration habits and preferences of academics were investigated via a pilot questionnaire (Zuccala et al., 2024), using Nahapiet and Ghoshal’s (1998) theory of social capital based on three related dimensions: cognitive, relational, and structural. Shared narratives and shared codes and languages are features of the cognitive dimension. A shared understanding of norms, feelings of interpersonal trust, feelings of mutual obligation, and a shared form of identification are all features of the relational dimension. Structural social capital, which is the third dimension, reflects ties to an appropriate organization (e.g. a university department/research field), ties (bonds/bridges) to others in a network, and the overall network configuration.

We use Social Cognitive Career Theory (Lent & Brown, 2019) as a theoretical foundation for considering how scientists prioritize collaborative relationships. At various points in a faculty member’s career, different pressures, expectations, and opportunities may shape how an individual strategizes their collaborative work. For example, pre-tenured faculty may focus more on preferences related to relational and cognitive networks, whilst senior faculty will seek to broaden their network, and thus alter its structure with the inclusion of more international, cognitive diverse relationships. Given this, we hypothesize that:

H1: Collaboration preferences and resulting structures change as academics advance through their careers.

A known challenge is that access to social capital varies in important ways specific to collaboration and career advancement. Elitism in science has created insurmountable boundaries for some individuals, as evidenced by hiring and other professional networks (Feeney & Bernal, 2010; ). Demographic differences in professional social networks and related social capital point to exclusion from resource-rich networks. For example, collaborative and other professional networks of underrepresented groups (women, people of color, LGBTQ) may start off with similarities at a cognitive and relational level, but a majority member network will change structurally more rapidly with the inclusion of international nodes, and nodes representing more cognitive diversity (Siciliano et al., 2018). Informed by studies of varied access to social capital, we hypothesize the following:

H2: Collaboration preferences of faculty across demographic groups will be similar, but underrepresented faculty members will experience greater difficulties with structural manifestation of these preferences.

Data and Methods Drawing from bibliometric and/or survey data, studies of social capital in academic science have used both global and ego-based network approaches (Kadushin, 2012) that capture various forms of professional networks. Rather than a dyadic co-author and collaborative approach, this study uses an approach similar to studies of scientific cosmopolitanism (Wang et al, 2019) or collaboration (Fox, 2010) where collaborators are identified by type or location. Because we want to examine the relationship between collaborative preferences and mobilized social capital, we use both survey and bibliometric data.

Our survey is based on the approach developed by Zuccala et al. (2024) where social capital was measured in terms of preferences. In this regard, social capital is not recognized as an outcome, but a catalyst or lubricant for cooperative work. In addition, we will match Scopus bibliographic data to survey respondents in order to examine which types of social capital preferences are linked with strategic outcomes.

These data constitute a significant part of a larger national survey (State of the Professoriate), implemented in October 2024 with faculty working in doctoral-serving universities across the United States. It was distributed to over 30,000 tenured and tenure-track faculty from 9 STEM disciplines across more than U.S. research-extensive institutions. The survey will close in January 2025, and given survey response trends we anticipate a 15% response rate.

Our analyses will begin with a descriptive analysis of how research collaboration preferences vary by demographic characteristics and career stage. Using bibliographic Scopus data we will further analyse respondents’ co-authorship network, where links represent ties to a shared article, and nodes represent a co-author by affiliation and gender (relational social capital) and area of expertise (cognitive social capital).

Preliminary Findings Although the survey is still in the field, we reviewed preliminary findings (1,000 respondents) in order to examine what we expect to see in the data. A difference of means analysis shows some differences in collaborative preferences by gender, rank and citizenship. Given the low proportion of non-whites in the U.S. professoriate, We will examine differences by race once all data are collected. These preliminary data are important in establishing patterns of collaboration preferences. The next step is to match them (and the full dataset) with publication data (already acquired and matched to individuals) to complete our analysis.

References: omitted due to length restriction.

ABSTRACT. 1. Introduction

The imperative to integrate academic research with societal needs has never been more pronounced than in the contemporary scientific landscape. Traditionally, the scientific paradigm has long prioritized pure inquiry or basic research (Stokes, 1997). In recent decades, given the evolving societal demands and expectations of publicly funded science, the U.S. National Science Foundation has paid special attention to the broader societal impact of research (NSF, 2022). Particularly pertinent to this funding shift is the notion of knowledge co-production, a practice that involves both academic and non-academic knowledge sources to better align scientific research with real-world applications (Adelle et al., 2020; Gerber et al., 2023).

Co-production of knowledge, defined as a collaborative, iterative approach integrating diverse perspectives of academics and non-academics throughout the research process, has become a cornerstone for sustainable and responsible research innovation (Dave et al., 2018). The inclusion of various actors in this process not only enhances the potential to produce actionable science but also fosters the development of solutions that are more applicable to addressing complex social problems (Mills et al., 2023; Norström et al., 2020). Engaging academic scientists in co-production therefore represents a pivotal shift towards addressing the multifaceted social problems to enhance the societal impacts of scientific research (Beckett et al., 2018; Zurba et al., 2022).

Despite its recognized value, knowledge co-production has not yet become a mainstream norm in the landscape of higher education institutions (Hart & Silka, 2020). The academic reward system, which typically prioritizes research excellence, often overlooks academic scientists’ outreach activities that directly collaborate with non-academics (Azman et al., 2019). As such, the engagement of academic scientists in knowledge co-production is not a common practice but is often rather driven by large science funding initiatives.

Yet, little is known about the specific factors influencing individual scientists’ perceptions of undertaking such non-traditional collaborative research endeavors. These factors can range from institutional support and cultures to individual motivations, values, skills, and/or the availability of resources. For instance, a study by Karlsdottir et al. (2023) highlights that individual factors such as age and academic discipline significantly influence perceptions of barriers to collaboration with industry partners, particularly those related to teaching obligations and university resources. This suggests that early career researchers and/or academics from certain disciplines might perceive greater obstacles to engaging in non-traditional collaborative activities such as co-production. Additionally, organizational culture and the broader institutional environment can also play a critical role in shaping individual perceptions and behaviors (Crow et al., 2020; Lattu & Cai, 2023). Understanding these factors is crucial as they can affect academic scientists’ propensity to engage in knowledge co-production, and our research seeks to fill this research gap. Therefore, drawing from the theory of planned behavior (Ajzen, 1991), our study will examine these research questions: 1) How do academic scientists perceive knowledge co-production and why? 2) How do academic scientists’ perceptions affect their intention to engage in knowledge co-production?

2. Data and Methods

To address the research questions, we will use data from a national survey (NETWISE III study) of over 30,000 tenured and tenure-track faculty from 9 STEM disciplines across U.S. Carnegie Classified R1 universities. The survey was launched in October 2024 and expected to be closed in January 2025. In this survey, respondents were asked about their perceptions and prior experiences regarding knowledge co-production with non-academics.

The dependent variables of our study are twofold: (1) the perception of research collaboration with non-academics, and (2) the intention to engage in knowledge co-production. As informed by the theory of planned behavior, the key constructs are operationalized through a series of survey questions that assess academic scientists’ attitudes toward research collaboration with non-academics, their perceived behavioral control/capabilities, subjective norms (e.g., research motivations and collaboration preferences), and intentions of engaging in knowledge co-production. Other variables include individual factors (e.g., gender, race/ethnicity, academic rank, and discipline), as well as organizational factors, such as departmental support, organizational culture, and the availability of resources for conducting collaborative research with non-academics.

Given that the proposed model includes both observed and latent variables, we will use the partial structural equation model (SEM) for data analysis. The SEM will provide robust estimations that allow us to test complex relationships among these variables simultaneously while accounting for the measurement error (Musil et al., 1998).

3. Preliminary Findings and Conclusion

Our preliminary analysis, based on over 900 survey responses collected so far, reveals that 60% of respondents have collaborated with a diverse range of non-academic partners for research purposes, including governments, nonprofits, community organizations, citizen science groups, etc. When asked about the proportion of their current research portfolio involving co-production, respondents reported an average of 38%. Moreover, these collaboration experiences span different phases of knowledge co-production. Notably, the most common activities include communicating research results to non-academics (44%) and co-authoring academic publications with non-academics (24%). Other collaborative efforts involve translating research findings into real-world applications (10%), and technology design and development (8%). Relatively smaller proportions of respondents reported collaborating with non-academics on other research tasks such as interpreting research results, data analysis, data access and collection, research design, and identifying research problems or questions.

Overall, our study findings will help identify the individual and organizational factors related to academic scientists’ propensity to engage in knowledge co-production. In particular, our research centers on individual scientists’ perceptions and experiences, offering a unique perspective by considering factors such as motivation and human capital. By shedding light on the motivation and barriers to academic scientists’ engagement in knowledge co-production, this study will contribute to the existing literature on this timely topic. Ultimately, this study will offer valuable insights for science policymakers and university administrators by informing them about research and university policies that can help facilitate knowledge co-production.

ABSTRACT. Isomorphism (or structural similarity) is a well-known concept both in population ecology theory (Hannan and Freeman, 1977, Aldrich, 1979) and neo-institutional organization theory (Meyer and Rowan, 1977, DiMaggio and Powell, 1983). Yet, only a handful of studies have attempted to operationalize and to empirically examine it (Croucher and Woelert, 2016, Woelert and Croucher, 2018, Seeber et al., 2014, Schofer and Meyer, 2005, Scott and Biag, 2016).

Isomorphism can be understood as the opposite of diversity, and one popular claim is that when organizations form an organizational field, their structural properties tend to become more similar over time. For example, there is considerable debate whether or not universities in Europe have become structurally more similar to each other in recent decades (Hüther and Krücken, 2016, Zapp et al., 2021).

Increased structural similarity may have detrimental consequences. For example, Hollingsworth (2006) argues that national research systems characterized by a high level of isomorphism have lower capabilities to produce scientific breakthroughs, whereas countries with low isomorphism levels produce more breakthrough science and technology. While Hollingsworth’s argument regarding different capabilities for breakthrough research has been empirically validated (Heinze et al., 2020), the pervasiveness of isomorphism in higher education systems has not been established for particular countries or regions. Therefore, this paper is an attempt to characterize isomorphism in one national higher education system over 25 years and thus create a basis for future comparative analyses.

For that purpose, the paper examines isomorphism quantitatively in German higher education, using the Relative Specialization Index. Drawing on a comprehensive data set that includes professorial staff, students, as well as basic and grant funding, the paper shows, first, that most variables show an isomorphic pattern, and second that public universities have become more isomorphic over time.

The paper is organized as follows. First, the paper reviews the literature on diversity and isomorphism in higher education and science studies. Second, it introduces an exemplary density function that displays how isomorphic distributions look like. After a short introduction of the dataset and some characterization of public universities in Germany, the paper presents empirical results, followed by a short discussion. Finally, the paper discusses future avenues for quantitative research on institutional isomorphism / diversity.

References

1. ALDRICH, H. E. (1979). Organizations and Environments, Englewood Cliffs: Prentice-Hall. 2. CROUCHER, G. & WOELERT, P. (2016). Institutional isomorphism and the creation of the unified national system of higher education in Australia: an empirical analysis. Higher Education, 71, 439-453. 3. DIMAGGIO, P. J. & POWELL, W. W. (1983). The Iron Cage Revisited - Institutional Isomorphism and Collective Rationality in Organizational Fields. American Sociological Review, 48, 147-160. 4. HANNAN, M. T. & FREEMAN, J. (1977). The Population Ecology of Organizations. American Journal of Sociology, 82, 929-964. 5. HEINZE, T., HEYDEN, M. V. D. & PITHAN, D. (2020). Institutional environments and breakthroughs in science. Comparison of France, Germany, the United Kingdom, and the United States. PLoS ONE, 15, e0239805. 6. HOLLINGSWORTH, J. R. (2006). A Path-Dependent Perspective on Institutional and Organizational Factors Shaping Major Scientific Discoveries. In J. T. HAGE & M. MEEUS (Eds.) Innovation, Science, and Institutional Change. (pp. 423-442). Oxford: Oxford University Press. 7. HÜTHER, O. & KRÜCKEN, G. (2016). Nested Organizational Fields: Isomorphism and Differentiation among European Universities. The University Under Pressure. (pp. 53-83). Emerald Group Publishing Limited. 8. MEYER, J. W. & ROWAN, B. (1977). Institutionalized Organizations: Formal Structure as Myth and Ceremony. The American Journal of Sociology, 83, 340-363. 9. SCHOFER, E. & MEYER, J. W. (2005). The worldwide expansion of higher education in the twentieth century. American Sociological Review, 70, 898-920. 10. SCOTT, W. R. & BIAG, M. (2016). The changing ecology of U.S. higher education: an organization field perspective. Research in the Sociology of Organizations, 46, 25-51. 11. SEEBER, M., LEPORI, B., MONTAUTI, M., ENDERS, J., DE BOER, H., WEYER, E., BLEIKLIE, I., HOPE, K., MICHELSEN, S., MATHISEN, G. N., FRØLICH, N., SCORDATO, L., STENSAKER, B., WAAGENE, E., DRAGSIC, Z., KRETEK, P., KRÜCKEN, G., MAGALHÃES, A., RIBEIRO, F. M., SOUSA, S., VEIGA, A., SANTIAGO, R., MARINI, G. & REALE, E. (2014). European Universities as Complete Organizations? Understanding Identity, Hierarchy and Rationality in Public Organizations. Public Management Review, 17, 1444–1474. 12. WOELERT, P. & CROUCHER, G. (2018). The Multiple Dynamics of Isomorphic Change: Australian Law Schools 1987-1996. Minerva, 56, 479-503. 13. ZAPP, M., MARQUES, M. & POWELL, J. J. W. (2021). Blurring the boundaries. University actorhood and institutional change in global higher education. Comparative Education, 57, 538-559.

ABSTRACT. Public research institutions face a growing demand to undertake research capable of addressing major socio-economic and environmental issues. The alignment between publicly funded research with societal challenges has been therefore at the core of the policy debate. The launch of the UN SDGs in 2015 reinforced these pressures on research organisations, in particular. Scholars in science policy and innovation studies have recently devoted considerable research efforts to examine the extent to which research aligns with funding targets (Aslan et al., 2024; Confraria & Wang, 2020; Wallace & Rafols, 2015) and societal needs (Ciarli & Ràfols, 2019; Confraria et al., 2024; Evans et al., 2014). These efforts have mostly focused on examining this alignment within research areas (for example, to what extent medical research efforts reflect the burden of diseases) and at the country level (for example, to what extent a country’s research aligns with SDGs that are most relevant to the country).

The organisational-level perspective has been, however, overlooked despite the pivotal role of organisations in research. Our understanding of the extent to which the directionality of universities’ research – one of the most prominent organisational actors in research systems – has shifted to address emerging societal challenges remains therefore limited. Universities are major knowledge producers and skills providers (Etzkowitz & Leydesdorff, 2000; Freeman, 1988; Pavitt, 1991). Recent evidence has also suggested that research has been increasingly ‘monopolised’ by universities as reflected in the global production of scientific publications where the involvement of private actors has become more selective (Arora et al., 2018; Larivière et al., 2018; Rotolo et al., 2022). As a result, a number of questions remain unaddressed: have the ‘best research skills’ been leveraged to undertake research capable of addressing societal challenges; how have inter-organisational collaborations shaped the alignment of research with societal challenges?; what is the role of private research organisations?

In this paper, we aim to fill this gap by examining the extent to which universities in the Italian Higher Education system have reoriented their research toward SDGs. In doing so, we examine the role of academic excellence and inter-organisational collaborations. First, we expect that SDGs can help universities build legitimacy in SDG-related research to attract additional resources and human capital (Mazzucato, 2018; Suddaby et al., 2017). Yet, we also expect that shifting research towards SDGs requires universities to reduce efforts in areas where they may excel. As a result, the tension between academic logic and societal benefits arises since research assessment is still mostly organised by disciplines with strong reputation incentives at the individual level. Second, we expect inter-organisational collaborations to be pivotal for universities to align their research with SDGs since the complexity of SDGs requires universities to go beyond the academic boundaries and those of a single discipline. Inter-organisational collaborations with private actors are likely to exert a positive effect on a university’s ability to undertake SDG-related research. Those collaborations provide transdisciplinary interactions, knowledge recombination opportunities with upstream/downstream domains (Fleming & Sorenson, 2001; Leahey et al., 2017), and access to the technological counterparts of scientific research (Barbieri et al., 2023). Collaborations with geographically distant actors are likely to exert a positive effect on a university’s ability to undertake SDG-related research. High-quality interactions are likely to happen over long distances and beyond the local networks (D’Este & Iammarino, 2010; Laursen et al., 2011).

Our empirical analysis examined the publication output of all universities in the Italian Higher Education system. We first queried Open Alex (Aria et al., 2024; Priem et al., 2022) to identify all publication records with Italian addresses. This led to an initial sample of over 2.6 million publications (2000-2023). We then matched the affiliation in publications with the list of 99 universities compiled by the Italian Ministry of Universities and Research – we also accounted for all research organizations affiliated with the universities (e.g. university hospitals). This led to a sample of over 1.9 publications involving at least one Italian university. We classified those publications by their SDG-related content by using keyword-based searches developed by (Jayabalasingham et al., 2019). About 22% of the publications in the sample could be related to SDGs. We then integrated these data with data from the Leiden ranking (data on universities’ share of highly-cited publications, share of collaborations with private actors, and share of short/long collaborations) and ETER (data on universities’ size in terms of student numbers). We estimated several panel regression models.

Our analysis suggested that Italian universities have increasingly aligned their research with SDGs. Yet there is considerable heterogeneity across SDGs – there is stronger alignment in the case of SDGs 3, 11, 12, 13, and 14. We also observe differences across universities where universities in the southern regions have shifted more towards SDGs than the universities in the northern regions. We also found that academic excellence exerts a negative impact on the shift towards SDG-related research. The more a university generates research of high academic impact (share of top 1% and top 5% highly cited publications) the less the university’s research directionality aligns with SDGs. Finally, the more a university collaborates with private organisations, the more its research directionality aligns with SDGs, while our model provided mixed evidence on the geography of collaborations.

These preliminary findings raise questions on whether the current academic incentive system hinders the use of the ‘best’ research skills to address SDGs as well as on the importance of universities engaging with the constellation of actors surrounding SDG-related challenges. While providing insights, at the organizational level, on the alignment of publicly funded research with societal needs, our study presents also some important limitations and opportunities for future research. Scholars have extensively debated the efficacy of various approaches to classify documents in terms of their SDG content (Kashnitsky et al., 2024). No approach has achieved broad consensus. At the same time, addressing the SDGs does not always require additional research. SDG challenges are embedded within existing socio-technical systems, necessitating systemic change rather than more research.

ABSTRACT. Introduction This paper presents the results of a study on the impact of the 2022 Russian full-scale invasion of Ukraine (the first invasion began in 2014) on the universities and other research institutions in 9 European countries bordering Ukraine, Russia, or Belarus.

Research Questions 1) What is the impact of the Russian invasion of Ukraine on the functioning of research and higher education institutions in countries bordering Ukraine, Belarus, or Russia? 2) Did the negative consequences of the Russian invasion of Ukraine have a more substantial impact on institutions with a relatively worse financial situation?

Methods & data A survey of scholars working in higher education and research institutions in Estonia, Finland, Hungary, Latvia, Lithuania, Moldova, Poland, Romania, and Slovakia. The survey questions were developed based on 11 individual interviews with scholars from the abovementioned countries.

Respondents were invited from a random sample of corresponding authors of publications indexed on the Web of Science database. The survey was conducted online (February- March 2024). It was available in English and 8 other languages used in the 9 countries covered by the study (Estonian, Finnish, Hungarian, Lithuanian, Latvian, Polish, Romanian, Slovak). We collected 3,743 fully completed surveys (response rate of 10%), with adequate representation of all surveyed countries (Estonia 334, Finland 407, Hungary 413, Latvia 486, Lithuania 527, Moldova 123, Poland 548, Romania 465, Slovakia 440). The data were analysed using descriptive statistics and logistic regression.

Results and conclusions The study uncovered a complex and multifaceted impact of the Russian invasion of Ukraine on higher education and research institutions across nine European countries bordering Ukraine, Russia, or Belarus. A significant proportion of respondents (22.5%) reported that their research activities were negatively or strongly negatively affected by the invasion. Key challenges included the suspension of scientific projects (9%), restricted access to research infrastructure (8.9%), and difficulties conducting field research (10.4%). Many scholars also experienced shifts in their academic focus, with 13.3% revising their research plans and 12.8% redirecting their research interests. Additionally, teaching activities were disrupted, as 10.6% of respondents indicated changes in their curricula due to the geopolitical turmoil.

The institutional response to the invasion was marked by significant strategic adjustments. Over 21% of respondents noted that their institutions had modified their development strategies in response to the crisis. Moreover, the invasion led to both opportunities and challenges in collaboration. While 17.4% of scholars reported new international collaboration opportunities, the invasion also disrupted existing partnerships, particularly with Russian institutions. The influx of Ukrainian students offered an unexpected benefit for many universities, partially mitigating the broader decline in student enrollment faced by the region. However, the cancellation of scientific events (18.2%) further underscored the invasion's pervasive impact on academic life.

The findings from logistic regression analysis reveal stark disparities in how institutions were affected. Financially weaker institutions faced more pronounced disruptions, reflecting and exacerbating pre-existing inequalities in the academic sector. Regression analysis indicated that institutions with stronger financial and reputational standing were better positioned to navigate the challenges posed by the crisis. Additionally, scholars deeply engaged in international collaboration—particularly with Ukrainian or Russian partners—were both more exposed to the invasion's adverse effects and more likely to capitalize on emerging opportunities for global engagement.

These results highlight broader systemic issues in the region’s higher education and research sector. Chronic underfunding, brain drain, and declining student admissions—already pressing challenges in many of the surveyed countries—were further intensified by the invasion. The unequal impact across institutions risks deepening these disparities, particularly for those in financially precarious situations. Yet, the crisis also spurred adaptive strategies, fostering new research collaborations and creating opportunities for institutional renewal.

The study underscores the need for targeted interventions to support vulnerable institutions and scholars. Strengthening international partnerships, increasing financial investment in underfunded universities, and promoting cross-border academic cooperation are essential to mitigate the invasion’s long-term repercussions. These measures not only address immediate challenges but also lay the groundwork for a more resilient and equitable higher education and research system in the European borderland. The findings offer insights for policymakers and academic leaders as they navigate the complex interplay of adversity and opportunity in a volatile geopolitical context.

ABSTRACT. The advancement of Artificial Intelligence (AI) technology has revolutionized traditional scientific paradigms, surpassing the cognitive and memory limitations of human scientists, and liberating researchers from the relentless tasks of literature review, data processing, and repetitive experimental procedures. Since the Japanese systems biologist Hiroaki Kitano proposed the use of AI for scientific discovery in 2016, discussions about the paradigm of intelligent science have never ceased, and AI is bound to fundamentally transform our entire research ecosystem. Currently, AI-powered scientific research has garnered significant attention across various disciplines, and discussions on the risks associated with AI applications have been initiated from multiple perspectives. However, few studies have systematically analyzed and elaborated on the risks associated with AI in scientific research from the perspective of the research process. To fill this gap and to better advance the development of the intelligent science paradigm, it is crucial to be vigilant against the potential risks that the AI-driven scientific revolution may bring, alongside enhancing research productivity. This paper utilizes a literature review approach, focusing on the application of AI in the scientific research process, and centers on the literature from the past five years. We systematically organizes the ethical and safety issues, data concerns, algorithm and model challenges, and cognitive risks associated with using AI in scientific research from both micro and macro perspectives. Furthermore, we analyzed global research progress and policy deployments to address these risks. Based on this foundation, this paper offers insights and recommendations for managing each of these risks. We sort out the use of AI tools in existing studies, hoping to answer the following questions: Q1: What are the potential risks when we apply AI to scientific research? Q2: What are the main studies that address these potential risks? Q3: What measures can be taken to mitigate these risks? It should be clear that this paper has a positive attitude towards the application of AI to the research process and revolutionizing the research paradigm. Our research aims to critically consider the intelligent science paradigm based on the existing literature, and to stimulate a broader discussion within the research community in order to better promote the development of this paradigm. The first part of this paper is a brief overview of the current status of AI application in the research process; the second part is an introduction to this research methodology and literature overview; the third part provides a detailed introduction to the risks of AI application in scientific research from four dimensions: data, algorithms and modeling, cognition, and ethics and security based on a review of the literature; and the fourth part offers some targeted recommendations for mitigating the aforementioned risks. This study divides the risks of AI application in scientific research into two main perspectives: the macro perspective of Ethics and Safety Risk (ESR for short) and the micro perspective of Algorithm and Model Risk (AMR for short), Data Risk (DR for short), and Cognitive Risk (ER for short). The micro perspective involves a granular examination of scientific research steps, scrutinizing the risks inherent in AI's application at each stage. Conversely, the macro perspective of ESR considers the entire scientific research process holistically, contemplating the broader implications of deeply integrating AI technology into the fabric of scientific inquiry and society at large. Data risk is mainly reflected in the unique attributes of the dataset. Algorithmic and model risk is mainly reflected in the use of the Big model for literature reading and knowledge extraction, the writing of research outputs, and the process of data analysis using AI algorithms. Cognitive risk occurs in the knowledge production and knowledge dissemination phases of the scientific research process, i.e., the analysis and writing of research results, and the publication and dissemination of research outputs. And then,this study provides a cursory review of current representative AI policy documents.From these policy documents, it is observed that the international community currently focuses on the overall framework-based governance of AI application risks, with most being encouragements and initiatives, and less attention is given to the application risks throughout the entire scientific research process. Therefore, based on the previous discussion on the application of AI in the entire scientific research process, this paper provides several targeted suggestions. It also needs to be emphasized that no single risk can be resolved solely by technology or by regulatory measures alone; a multifaceted effort is required to promote the scientific deployment and reasonable operation of the intelligent scientific paradigm.

ABSTRACT. Introduction

Large language models (LLMs) are increasingly being integrated into scientific research to automate and enhance various tasks, from code generation and data analysis to literature review, content creation, and even scientific writing (Lin., 2023). However, studies consistently show that LLMs carry biases in gender, race, and language, among others (Gallegos et al., 2024). Within the scientific context, known disparities in prestige, gender, and race (Wapman et al., 2022; Kozlowski et al., 2022) could contribute to these issues. However, the presence and impact of these biases in the scientific domain remain underexplored. In this study, we aim to detect and examine potential disparities in LLMs for and about science, focusing on gender, race, location, field of study, socioeconomic status, and academic prestige. We analyze their impact on the representation of scientific knowledge and researchers. We apply a well-known method for estimating biases, the Word Embeddings Association Test (WEAT), to measure issues in popular text models, BERT and SciBERT. We also analyze biases directly in an advanced LLM model, Anthropic’s Claude 3.5 Sonnet. Our results reveal significant biases across all dimensions examined, with a consistent tendency to favor dominant groups, such as people and institutions in the Global North.

Background Previous studies show that LLMs are increasingly integrated into scientific research, enhancing research productivity, diversity, and innovation by leveraging their three key attributes: intelligence, versatility, and collaboration (Lin., 2023). Although there is limited research on biases in LLMs within scientific contexts, prior studies have documented significant disparities in the research workflow related to prestige, gender, and race. Prestigious institutions dominate faculty placements, reinforcing hierarchical structures (Way et al., 2019). Gender disparities and career sacrifices disproportionately affect women, while racial disparities hinder the career progression of faculty of color (White et al., 2021).

Data and Methods Our analysis focuses on two main investigations: 1) whether disparities exist in underlying models used by LLMs, and 2) whether LLMs’ responses about science favor certain geographical or demographic groups. The first investigation uses the embedding level of LLMs. We apply the WEAT method, which calculates semantic distances between target and attribute word pairs using cosine similarity, to detect associations that reveal biases. This analysis is performed in BERT, which is general, and SciBERT, which is trained on scientific literature. In our experiments, we studied attribute and target words. Attribute words represent two sets of positive and negative words that we want to contrast with target words such as populations. We use positive and negative words, respectively, generated by GPT-4 to describe five key areas of research and academia including performance, productivity, teaching effectiveness, funding acquisition, and awards/ recognition. A set of targets was obtained from Caliskan et al. (2017), representing European vs. African-American names, male vs. female names, and science vs. art terms. We used GPT-4 to generate Global North vs. South, high vs low socioeconomic status, and prestigious vs. non-prestigious institutions. The second investigation involves LLM prompting tasks. The first task simulated email inquiries to professors about Ph.D. positions, varying the gender and location of the sender. Responses were rated on a Likert-like scale from 1 (Highly unlikely to reply) to 5 (Highly likely to reply). In another task, we simulate manuscript cover letter submissions, with the editor’s decision to desk reject or send for peer review.

Results The first investigation revealed consistent biases in both BERT and SciBERT through the WEAT task. SciBERT shows stronger biases in associating the overall dominant groups (eg. male and Global North names) with research and award successes than BERT (0.641 vs 0.569 and 0.824 vs 0.424 respectively on a scale from +2, indicating a strong bias toward dominant groups, to -2, toward underrepresented groups), and specifically, the Global North institutions with teaching success and research productivity (0.849 vs 0.670 and 1.199 vs 0.574 respectively), while BERT more strongly links Northern and European traits with academic success in teaching (1.232) and research productivity (1.226). The second investigation also showed consistent biases in both email inquiry and cover letter tasks. The email response task revealed disparities across gender and regional origin. Female candidates received a higher likelihood of response (M = 2.62, SEM = 0.12) compared to male candidates (M = 2.50, SEM = 0.11), while Global North candidates received significantly higher response rates (M = 2.90, SEM = 0.15) than Global South candidates (M = 2.22, SEM = 0.13). In the editorial review task, female authors from the Global North had the highest likelihood of their manuscript being reviewed (M = 3.34, SEM = 0.04), followed by female authors from the Global South (M = 3.07, SEM = 0.05).

Conclusion Our results emphasize the potential for LLMs to amplify existing biases, even as they become more integrated into scientific workflows (Wagner et al., 2022). Future research will expand prompt categories and test models like Llama 3 and ChatGPT, examining bias persistence across platforms. This will build on existing research on LLM biases, informing more equitable use of AI within scientific and educational contexts.

References Gallegos et al. (2024). Bias and fairness in large language models: A survey. Computational Linguistics, 1-79. Caliskan et al. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183-186. Clauset et al. (2015). Systematic inequality and hierarchy in faculty hiring networks. Science Advances, 1(1), e1400005. Wapman et al. (2022). Quantifying hierarchy and dynamics in US faculty hiring and retention. Nature, 610(7930), 120-127. Kozlowski et al. (2022). Intersectional inequalities in science. Proceedings of the National Academy of Sciences, 119(2), e2113067119. White et al. (2021). Race, gender, and scholarly impact: Disparities for women and faculty of color in clinical psychology. Journal of Clinical Psychology, 77(1), 78-89. Way et al. (2019). Productivity, prominence, and the effects of academic environment. Proceedings of the National Academy of Sciences, 116(22), 10729-10733. Lin, Z. (2023). Why and how to embrace AI such as ChatGPT in your academic life. Royal Society open science, 10(8), 230658.

ABSTRACT. Developing Data-Centric Approaches for Assessing the Risk of Research Sponsor Bias at Scale

Funders impact the knowledge produced (Gläser & Velarde, 2018). Particularly, industry sponsorship can affect the outcomes of research (Lundh et al., 2018). Consequently, managing financial conflicts of interest has become important in clinical medicine (Torgerson et al., 2022). However, determining in the aggregate who funded research on a given topic, and the type of research funder, is currently difficult. Assessing the extent and nature of industry funding is particularly important for assessing the risk of bias in research. Scholars in agnotology describe how "socially constructed ignorance" such as "manufactured certainty" (Leefmann, 2021; Pinto, 2017) has resulted from industry tactics (Goldberg & Vandenberg, 2021) such as tobacco industry lawyers’s involvement in selecting the research to be funded externally (Bero et al., 1995) as well as within the company (Hanauer et al., 1995). Philosophers of science have argued that a scholarly community can become biased without any individual being biased, a phenomenon they refer to as "Experimentation by industrial selection" (Holman & Bruner, 2017). Funding acknowledgement data is becoming available from sources such as Crossref, Dimensions, Lens, Scopus, and Web of Science (Kramer & de Jonge, 2022), with varying data quality (Álvarez-Bornstein et al., 2017; Kramer & de Jonge, 2022; Liu, 2020; Liu et al., 2020). This data is not yet comprehensive: in Crossref, only about 14% of data of records on COVID-19 research had a funder identifier (Mugabushaka et al., 2022). And, authors’ financial disclosures are likely incomplete. Yet we can start to envision data-driven approaches to assessing the risk of sponsor bias in research at scale. We have developed a prototype digital library application "WhoFundedIt" that uses funding information from Crossref to visualizes who funds research on a specific topic, given any set of DOIs. We are developing a series of case studies using bibliographic data, drawing on cases of well-documented sponsor bias on topics such as alcohol, gambling, pharmaceuticals, sugar, and tobacco (Legg et al., 2021). By comparing topics with known sponsor bias to non-controversial, non-applied topics (which we assume have less risk of sponsor bias) we seek to to identify risk signatures that "WhoFundedIt" could incorporate in the future.

Acknowlegements NSF 2046454 CAREER: Using network analysis to assess confidence in research synthesis. 2024–2025 Perrin Moorhead Grayson and Bruns Grayson Fellow, Harvard Radcliffe Institute for Advanced Study. Thanks to Theodore Dreyfus Ledford for research on conflict of interest; Saish Desai, Johan Krause, and Malik Salami for assessing existing tools for extracting funder names from XML; Shashank Kambhatla, Hannah Smith, and Deyuan (Doreen) Yang for work on the WhoFundedIt prototype, and members of our Citizen’s Advisory Board.

References Álvarez-Bornstein, B., Morillo, F., & Bordons, M. (2017). Funding acknowledgments in the Web of Science: Completeness and accuracy of collected data. Scientometrics, 112(3), 1793–1812. https://doi.org/10.1007/s11192-017-2453-4 Bero, L., Barnes, D. E., Hanauer, P., Slade, J., & Glantz, S. A. (1995). Lawyer control of the tobacco industry’s external research program. The Brown and Williamson documents. JAMA, 274(3), 241–247. Gläser, J., & Velarde, K. S. (2018). Changing funding arrangements and the production of scientific knowledge: Introduction to the special issue. Minerva, 56(1), 1–10. https://doi.org/10.1007/s11024-018-9344-6 Goldberg, R. F., & Vandenberg, L. N. (2021). The science of spin: Targeted strategies to manufacture doubt with detrimental effects on environmental and public health. Environmental Health, 20(1), 33. https://doi.org/10.1186/s12940-021-00723-0 Hanauer, P., Slade, J., Barnes, D. E., Bero, L., & Glantz, S. A. (1995). Lawyer control of internal scientific research to protect against products liability lawsuits: The Brown and Williamson Documents. JAMA, 274(3), 234–240. https://doi.org/10.1001/jama.1995.03530030054034 Holman, B., & Bruner, J. (2017). Experimentation by industrial selection. Philosophy of Science, 84(5), 1008–1019. https://doi.org/10.1086/694037 Kramer, B., & de Jonge, H. (2022). The availability and completeness of open funder metadata: Case study for publications funded by the Dutch Research Council. Quantitative Science Studies, 3(3), 583–599. https://doi.org/10.1162/qss_a_00210 Leefmann, J. (2021). How to assess the epistemic wrongness of sponsorship bias? The case of manufactured certainty. Frontiers in Research Metrics and Analytics, 6. https://doi.org/10.3389/frma.2021.599909 Liu, W. (2020). Accuracy of funding information in Scopus: A comparative case study. Scientometrics, 124(1), 803–811. https://doi.org/10.1007/s11192-020-03458-w Liu, W., Tang, L., & Hu, G. (2020). Funding information in Web of Science: An updated overview. Scientometrics, 122(3), 1509–1524. https://doi.org/10.1007/s11192-020-03362-3 Lundh, A., Lexchin, J., Mintzes, B., Schroll, J. B., & Bero, L. (2018). Industry sponsorship and research outcome: Systematic review with meta-analysis. Intensive Care Medicine, 44(10), 1603–1612. https://doi.org/10.1007/s00134-018-5293-7 Mugabushaka, A.-M., van Eck, N. J., & Waltman, L. (2022). Funding COVID-19 research: Insights from an exploratory analysis using open data infrastructures. Quantitative Science Studies, 3(3), 560–582. https://doi.org/10.1162/qss_a_00212 Pinto, M. F. (2017). To know or better not to: Agnotology and the social construction of ignorance in commercially driven research. Science & Technology Studies, 30(2), Article 2. https://doi.org/10.23987/sts.61030 Torgerson, T., Wayant, C., Cosgrove, L., Akl, E. A., Checketts, J., Dal Re, R., Gill, J., Grover, S. C., Khan, N., Khan, R., Marušić, A., McCoy, M. S., Mitchell, A., Prasad, V., & Vassar, M. (2022). Ten years later: A review of the US 2009 Institute of Medicine report on conflicts of interest and solutions for further reform. BMJ Evidence-Based Medicine, 27(1), 46–54. https://doi.org/10.1136/bmjebm-2020-111503

ABSTRACT. The National Center for Science and Engineering Statistics (NCSES) within the U.S. National Science Foundation (NSF) is the principal source of analytical and statistical reports, data, and related publications that describe and provide insight into the nation’s science and engineering resources. NCSES collects and provides data on the science and engineering workforce; research and development (R&D); competitiveness in science, engineering, technology, and R&D; and the condition and progress of STEM education in the U.S. These data are collected as sample surveys or census.

The CHIPS and Science Act, PL 117-162, was signed into law in August 2022. Section 10375 establishes a National Secure Data Service Demonstration Project (NSDS-D) to "develop, refine, and test models to inform the full implementation of the Commission on Evidence-Based Policymaking recommendation for a governmentwide data linkage and access infrastructure for statistical activities conducted for statistical purposes, as defined in chapter 35 of title 44, United States Code." The Advisory Committee on Data for Evidence Building: Year 2 Report recommends several approaches for the NSDS that focus on supporting a high-quality user experience. Recommendation 3.4 states "The NSDS should employ data concierges to help users refine their research projects, discover relevant data, and acquire access to that data.”

Statistical agencies produce a large amount of data and analysis every year. These data can answer questions such as “How do most people die?” and “How many employed engineers with doctorates live in Arkansas?”. However, simply typing these questions into an internet-wide search engine or search bar on a statistical agency website will not always return the direct information requested. In addition, it may be difficult to know with an internet-wide search if the result is based on a statistic from federally published data or other non-federal data. The expansion of machine learning, including natural language processing and AI, may provide an updated approach to directly answering user queries related to public statistical information. The team is working to build a pilot tool to respond to users’ queries using publicly available data on statistical websites.

In the first part of the pilot, the team is building a Retrieval Augmented Generation (RAG) based system (process that controls the output of the Large Language Model before generating a response by leveraging database access) that is compatible with and builds on the open-source framework behind Google’s data commons. The chatbot will focus on three data sets that represent unique ways statistical agencies publish their data: (1) public use file, (2) data tables, and (3) analytical reports.

An important feature of this effort is how users interact with the system. The pilot tool, grounded in the Data Commons framework, offers: • Built-in visualizations, including timeline charts, scatter plots, and maps. • A natural-language query interface • Interactive and programmatic query capabilities. • Exploration tools for interacting with extensive datasets. • A Statistical Variable Explorer for exploring data variables. • A Knowledge Graph browser for examining data relationships. • Customization capabilities for organization-specific data implementations. • An enhanced Explore interface that uses large language models (LLMs) to map user queries to public datasets, generating relevant visualizations.

These features are designed to make public data more accessible, useful, and relevant for a broad range of users, including those in science, policy, journalism, and more.

This research presentation will introduce the pilot tool to the audience and detail salient technical details and lessons learned. Training of the model will be presented along with quality metrics such as precision and recall and benchmarking tests. In addition, we will present lessons learned about the size of input data tables, making statistical data “AI ready”, and other engineering issues encountered while building the pilot tool.

ABSTRACT. 1. Introduction In the globalised and competitive academic landscape, universities strive to enhance their research excellence, reputation, and visibility. International highly cited researchers (HCRs), are known to significantly contribute to a university’s scientific advancements, increase institutional prestige, and foster international collaborations. International HCRs not only elevate a university's research output but also play a pivotal role in securing funding, improving global rankings, and enriching the academic environment. Academic networks are known to shape research productivity and mobility. A university's position within international research networks can influence its visibility and accessibility to HCRs. Closeness centrality is a fundamental concept that measures how close a node is to all other nodes in a network. In the organisational knowledge network, the concept captures how efficiently a university can utilise knowledge within the network. High closeness centrality in the international HCRs network implies that a university can quickly reach and be reached by other prominent international research organisations and HCRs globally. Hence, making them more attractive destinations for international HCRs seeking to maximise their research impact and collaborative opportunities. Therefore, we propose: H1. A university’s closeness centrality in the international HCR network is positively correlated with its share of incoming international HCRs. Furthermore, the average shortest path reflects how quickly and efficiently information can flow through the network from one actor to another. A lower average shortest path indicates that the organisation is closely connected to others, facilitating rapid exchange of knowledge amongst organisations and researchers globally. It also reduces coordination costs and barriers to collaboration, as researchers can connect with fewer intermediaries. From the perspective of international HCRs, universities with a lower average shortest path length offer a more efficient platform for global collaboration and knowledge exchange. These organisations provide access to a broad network with minimal effort, making them attractive for HCRs aiming to maximise their impact and productivity. Thus, we hypothesise: H2. A university’s average shortest path in the international HCR network is negatively correlated with its share of incoming international HCRs. Similarly, the local clustering coefficient measures the likelihood that a node's neighbours are also connected to each other, forming tightly knit groups within the network. A high local clustering coefficient indicates strong interconnections among an organisation's collaborators, leading to cohesive subgroups. While this can foster trust and facilitate the sharing of tacit knowledge within the cluster, it may also result in redundancy of information and limit exposure to diverse perspectives. High clustering can create echo chambers where novel ideas are less likely to emerge due to the homogenisation of information. On the other hand, a lower local clustering coefficient indicates that a university's collaborators are less interconnected, exposing the university to a wider array of information from different parts of the network.For international HCRs, universities with lower local clustering coefficients may be more appealing because they offer access to diverse networks and reduce the risk of redundant collaborations. These universities provide opportunities to connect with unconnected or loosely connected collaborators, enhancing the potential for impactful research. Hence, we propose: H3. A university’s local clustering coefficient in the international HCR network is negatively correlated with its share of incoming international HCRs. 2. Method and Results We utilised panel data comprising 86 German universities from 2005 to 2020. Our dependent variable is the number of incoming international HCRs to a given university as a share of its total publishing authors for a given year. We use the author’s change of affiliations to determine their joining to a university. Our key explanatory variables are network measures derived from the universities' HCR collaboration networks. The nodes of our networks are the universities, and the edges are the number of publications among a university and other publishing organisations. Our network measures are closeness centrality, average local shortest paths, and local clustering. To establish whether the effect is significantly different when establishing connections domestically or internationally, each measure is calculated for a university’s international and domestic network. A university’s international network includes the connections amongst the focal domestic university and other international organisation, whilst the domestic network is between domestic organisations. We included controls for time, disciplines, regions, and university characteristics. We ran three separate regressions for each of the network measures using fixed effects. In the first model, the coefficient for external closeness centrality is negative and statistically significant (-0.0124***), while its interaction is positive and significant (0.0138***). This indicates a U-shaped relationship between external closeness centrality and the share of incoming international HCRs. In the second model, the external average shortest path length has a negative coefficient (-0.0013*) with a positive squared term (0.0002*), also suggesting a U-shaped relationship. Whilst, in the third model, the international local clustering coefficient shows a negative effect (-0.0030**) with a positive squared term (0.0037**), confirming the U-shaped relationship. We find domestic network measures are not significant throughout. 3. Discussion The results highlight the non-linearity between the international network measures and the share of incoming international HCRs, which contrasts with our initial hypotheses that posited linear relationships. One possible explanation is that, on the one hand, universities in peripheral positions offer unique research opportunities that attract international HCRs seeking niche areas. On the other hand, well-connected universities become hubs, providing extensive collaboration opportunities, which are attractive to international HCRs. Universities in the middle range might not offer the advantages of peripheral or well-connected organisations, making them less attractive to international HCRs. Hence, high connectivity may appeal to those seeking strong, tight-knit research communities with established collaboration patterns. Conversely, low connectivity may attract HCRs interested in bridging structural holes and accessing diverse, unconnected networks. These non-linear relationships challenge the simplistic assumption that better network positions attract more HCRs. Instead, they highlight the complexity of network dynamics and the varied preferences of HCRs. Moreover, the insignificant effects of domestic network measures indicate that domestic collaborations within Germany do not impact the attraction of international HCRs. This emphasises the importance of international over domestic networking in the global competition for top researchers.

ABSTRACT. Introduction

Across academic fields in the U.S., tenured and tenure-track faculty are approximately 25 times more likely to have a parent with a PhD than the typical American, and this tendency is even higher at elite universities [1]. The impact of this lack of socioeconomic diversity on scientific discoveries and innovations remains largely unknown. While gender, race, and ethnicity are known to influence the researcher's specialty (see [2]), socioeconomic background has not been studied due to the lack of data. This project studies and uncovers large-scale differences in scholarship across socioeconomic status (SES). To do so, we apply latent Dirichlet allocation (LDA), an unsupervised topic model, to publications of U.S. tenured and tenure-track faculty for which we have self-reported SES. Our results show that some research topics exhibit strong correlations with the SES of faculty.

Background and Research Question

Previous research shows that the composition of the scientific workforce shapes the pace and direction of discovery, and that increased diversity can produce more innovative science. Hence, representational disparities in the scientific workforce can represent a kind of “epistemic inefficiency” in the production of knowledge, as too few individuals focus on some questions and too many on others. Prior work has linked research topics with researcher gender and to a lesser extent to race and ethnicity. This systematic association suggests losses of innovation in identity-correlated topics like women’s health, ethnic studies, history, and heritable or infectious diseases. No comparable research exists for SES, largely due to the inability to infer SES algorithmically from bibliographic, name, or employment data, and because SES is complex and multidimensional. Instead, SES must be measured through self-reporting, which is difficult to perform at scale.

Studies have shown that childhood SES influences individuals’ educational attainment, including obtaining a PhD, academic socialization, and building ‘doctoral capital’. This literature illustrates the multiplicity of pathways by which SES can influence a person’s scholarly trajectory. Our study extends this to characterize the SES influence on scientific discovery and innovation. Understanding how researcher SES correlates with different scholarly outputs raises several questions, including: What scientific discoveries and innovations are we missing because the scientific workforce is dominated by higher SES researchers? How does SES’s impact on scholarship intersect with researcher gender, race, and ethnicity?

Data and Methods

Our study links three sources of data with state-of-the-art methods to address these questions: (1) a census of tenured and tenure-track faculty at U.S. PhD-granting universities from the Academic Analytics Research Center (AARC); (2) a complementary survey of faculty that collects self-reported SES; and (3) clean publication records from the open-source bibliometric database OpenAlex. Our census encompasses 239,949 tenure-track or tenured faculty across 111 academic fields at PhD-granting institutions in the U.S. who were active in 2011 to 2020. Here, we present results from combining self-reported SES variables for faculty in Computer Science (CS) and Biology, for whom we algorithmically linked with OpenAlex researcher IDs using their names and then affiliation institutions using string functions. Given the data available from the survey, we use the highest level of parental education reported by the respondent as a proxy for childhood SES, acknowledging the limitations of only using one of three dimensions commonly used to estimate SES. After cleaning the resulting 3-way linked dataset (faculty, SES, and publications) to correct incorrect linkages, we concatenated the publication abstract, title, and venue and applied Latent Dirichlet Allocation (LDA) topic modeling to represent each faculty member’s record of scholarship as a k-dimensional vector.

In the fields of Computer Science (CS) and Biology, the corpus has 40,978 and 162,190 publications respectively, for 363 computer scientists and 1,388 biologists. When fitting the LDA model on each field and stratifying by the researcher SES, we can observe if topic weights are correlated with SES. For each field, we estimated the SES gradient with a linear regression.

Results

Our analysis reveals that some research topics within a field exhibit strong correlations with faculty’s SES, while others do not. For instance, relative to lowest-SES researchers in their field, we find that the highest-SES individuals in CS are 1.47 times more likely to study “Network Systems” and in Biology are 1.23 times more likely to study “Bioinformatics.” For these two subtopics, we see a clear and positive correlation between topic proportion and SES level. Conversely, we find little correlation with SES in “Machine Learning and AI” for CS (odds: 0.97) or “Biochemistry” for Biology (odds: 0.98). The topics correlated with higher SES background –“Bioinformatics” and “Network Systems”– are technical and resource-intensive subfields. This pattern is consistent with an association between more technical fields and more privileged socioeconomic backgrounds. However, the “Machine Learning and AI” subfield in CS is also highly technical, but shows no correlation with SES background, suggesting a more complex relationship. Our further research agenda will seek to clarify this.

Discussion

These initial results provide evidence for our hypothesis that SES background can and does shape scholarship choices, even among highly trained researchers and within STEM fields like CS and Biology, which have fewer subfields that prioritize identity-focused scholarship. These also complement existing literature that links gender, race, and ethnicity to the researcher's specialty (see [2]). Characterizing how researchers’ SES shapes what scholarship they pursue or how they pursue it is especially important given the lack of socioeconomic diversity within the scientific workforce. In future work, we will compare these effects across our 111 sampled fields, shedding new light on how the composition of the scientific workforce shapes scientific discovery and innovation. Finally, understanding the concentration of researchers by SES across subfields and topics in academia will aid policies that tackle inequalities in academia, particularly as research has shown that the topic of choice can influence scholarship [3].

References

[1] Morgan, A. C. et al. Socioeconomic roots of academic faculty. Nat. Hum. Behav 6, 1625–1633 (2022).

[2] Kozlowski, D. et al. Intersectional inequalities in science. PNAS USA 119 (2022).

[3] Hoppe, T. A. et al. Topic choice contributes to the lower rate of NIH awards to African-American/black scientists. Science Advances 5 (2019).

ABSTRACT. Background Academic hiring outcomes [1] and scholarly productivity [2] are deeply rooted in the prestige of a professor’s PhD institution and faculty institution, but little is known about the paths tenure-track professors took before their PhDs. For instance, while the current professoriate is 20 times more likely to have a parent with a PhD than the U.S. population [3], other stages in the preparation and background of today’s faculty remain poorly understood, including their undergraduate education. Given how important prestigious undergraduate institutions are for granting students access to elite jobs in general [4, 5], this raises the question of whether and how a professor’s undergraduate pedigree impacts their future career.

Data & Methods To investigate the undergraduate experiences of faculty, we drew together three complementary datasets. We compiled a detailed dataset of the educational trajectories of all 16,006 tenure-track faculty active in 2011 and representing all career stages from 404 PhD-granting Computer Science, Business and History departments in the U.S., who attended 960 U.S. undergraduate institutions. We combined this person-level data with institutional metadata from the U.S. College Scorecard and with institutional prestige rankings from the U.S. News & World Report for both liberal arts colleges and universities. Together, these datasets allowed us to compare the undergraduate education of professors to that of the general U.S. undergraduate population, and the undergraduate education of elite professors to non-elite professors.

To evaluate institutional relationships in faculty trajectories, we constructed a career trajectory network, defining each node as an institutional transition—representing either the transition from undergraduate to PhD or the transition from PhD to professorship—with directional links.

Results Compared to U.S. undergraduate students at 4-year institutions, U.S.-educated professors are more likely to have attended private, research-intensive, and elite universities. Nearly 70% of U.S. undergraduates attend public universities, compared to only 47% of U.S.-educated professors, and roughly half of U.S. undergraduates attend research-intensive universities, compared to almost all (91%) U.S.-educated professors.

In fact, the trajectories of many faculty pass through only a handful of highly elite universities: one third of faculty in our sample attended one of 15 schools for their undergraduate degree, while only one in 100 U.S. undergraduate students attends one of those schools. Further still, nearly one out of every six professors attended one of just 5 schools for their undergraduate degree: Harvard, MIT, Yale, University of California-Berkeley, or Cornell, versus just four in 1000 U.S. undergraduates. Interestingly, not only are large proportions of U.S. faculty also trained outside the U.S., but these proportions vary markedly by field (51% in Computer Science versus just 17% in History).

Even compared to each other, elite professors are more likely to have attended elite undergraduate institutions than non-elite professors: 38% of professors working at a top-5 school received their undergraduate degree from a top-5 school. Similarly, 41% of professors who received their PhD from a top-5 school in their degree program also received their undergraduate degree from a top-5 school. While there is no single path to prestigious faculty positions, the odds become much higher with a more prestigious pedigree—the conditional probability of becoming a tenure-track professor at an institution ranked in the top ten percent (given that they are a professor) is 32% for a professor who attended an undergraduate institution ranked in the top ten percent, versus just 6% for a professor who attended an undergraduate institution ranked in the lowest ten percent.

Further, our network analysis shows that elite undergraduate institutions gave faculty a relative prestige boost in where they got a PhD and faculty job. Faculty who graduated from undergraduate institutions in the top twenty percent were more likely to “move up” the hierarchy—they attended PhD institutions and were hired as faculty by institutions more prestigious than their undergraduate institution. However, faculty who graduated from undergraduate institutions in the bottom twenty percent were more likely to “move down” the hierarchy—they attended PhD institutions and were hired as faculty by institutions less prestigious than their undergraduate institution.

In order to identify the steepest inequalities in trajectories in the network, we looked at the largest strongly connected component, the largest sub-network within the network where there is a path between every pair of nodes within the sub-network. While we found that only a small fraction of highly prestigious institutions (CS: 0.5%; Business: 1.5%; History: 5%) exchange graduates among each other, this small proportion of institutions decreases even further in randomized networks, suggesting that a specific preference among the most prestigious institutions exists for PhD students and faculty from highly prestigious undergraduate institutions.

Significance This work contributes to a growing literature focused on understanding barriers to scientific jobs, particularly in light of ongoing efforts to diversify academia, including by socioeconomic background. We find there is a distinct pipeline from prestigious undergraduate institutions, to prestigious graduate school programs, to prestigious faculty jobs. The cumulative advantage offered by an elite undergraduate degree, which is most often determined by social class and not ability [5, 6], reproduces a homogenous group of tenure-track professors trained at a few highly prestigious institutions. These findings set the stage for future investigations aimed at increasing access to who becomes a professor, and will be valuable to PhD admissions committees and faculty hiring committees who want to make more equitable admissions and hiring decisions.

References [1] Clauset, Arbesman & Larremore. “Systematic inequality and hierarchy in faculty hiring networks.” Science Advances, 1(1):e1400005, 2015. [2] Way et al. “Productivity, prominence, and the effects of academic environment.” PNAS, 2019. [3] Morgan et al. “Socioeconomic roots of academic faculty.” 2021. [4] Rivera. Pedigree: How Elite Students Get Elite Jobs. Princeton University Press, 2016. [5] Chetty et al. “Mobility report cards: The role of colleges in intergenerational mobility” (No. w23618). National Bureau of Economic Research, 2017. [6] Hoxby & Avery. “The missing one-offs: The hidden supply of high-achieving, low income students” (No. w18586). National Bureau of Economic Research, 2012.

ABSTRACT. Acknowledging the US federal government’s research support in patents is formally required by law when patents originate from the US government’s research funding. According to 35 United States Code section 200-202 (35 U.S.C. § 200-202, as known as the “Bayh-Dole Act”), recipients or contractors of the federal government’s research grants are allowed to own resulting patents. Yet, they are also obligated to acknowledge the government’s support in patents to entitle the federal government to the march-in right, which enables the government to require patent owners to “grant a nonexclusive, partially exclusive, or exclusive license to a responsible applicant or applicants with reasonable terms when the original patent holder fails to take effective steps to achieve practical application of the invention, or requirements to alleviate health or safety needs that are not reasonably satisfied by the patent holder.” Even if the patent owner refuses to do so, the government can grant the license to the third party by itself (see, 35 U.S.C. §203(a)). In addition to research funding, the US federal government supports research by allowing federal-government-affiliated researchers (hereafter, federal researchers) to collaborate with scientists or engineers in other organizations. For this type of support, the government’s contribution to resulting patents can be acknowledged by including federal researchers as inventors.

Although federal research support plays a crucial role in driving innovation, its contribution could be undermined when the US federal government’s research support in patents is not acknowledged. On the one hand, because the federal government is entitled to exercise patents arising from its research support for public safety or health when the government’s involvement in patents is properly acknowledged, failure to document the government’s research support in patents limits potential social benefits that a patented invention can realize through governmental use. On the other hand, not acknowledging the government’s research support in resulting patents undermines the government’s ability to promote the use of research outcomes for subsequent commercial development or encourage market competition. These concerns become even more compelling considering that private stakes lying in the exclusive use of the research outcomes may lead patent holders not to acknowledge the government’s involvement.

To what extent is the government’s research support acknowledged in patents? Are private stakes in the exclusive use of research outcomes related to underappreciation of the government’s support in resulting patents? In the present study, I address this gap by analyzing about 84,000 US patent-paper pairs (PPPs) on research outcomes arising from the US federal government’s research support. Research outcomes can be published through academic research papers as well as patent applications. I defined a patent and research paper as a “pair” if they were on “equivalent” scientific findings while sharing at least one common individual among the authors of the paper and inventors of the patent. In my analysis, I focused on patents paired with research papers that acknowledged the US government’s research support. My approach was based on the rationale that paired patents to research papers acknowledging the federal government’s research support should have consistently credited the involvement of the government therein.

Overall, of patents on research outcomes originating from federal research support, 28% did not acknowledge the government’s research support. The number of patents acknowledging the government’s involvement in patents by funding organizations at the department level demonstrated that the underappreciated governmental research support in patents was pervasive across the body of the federal government providing research support.

I further examined whether three variables capturing the prominence of private stakes in the use of research outcomes—for-profit orientation of patent owner, presence of research support by private entities, and commercial potential of inventions—were negatively correlated with the likelihood that the federal government’s research support was acknowledged in patents. First, given the very nature of firms as the entity maximizing private return to investments in their research, I investigated if patents assigned to firms exhibited a significantly lower likelihood of acknowledging the government’s research supports in patents than patents assigned to other types of entities. My data demonstrated that this was the case. On average, 48% of patents assigned to firms acknowledged the federal government’s support. In contrast, more than 68% of patents that were assigned to other types of entities acknowledged the government’s research support.

Second, considering the presence of private research funding or the involvement of firm researchers (hereafter, private research support) as the indicator of revealed prominence of private stakes in the use of research outcomes, I tested if the presence of private research supports decreased the likelihood of acknowledging government’s research support in resulting patents. It was found that 55% of patents arising from private research support along with the US government’s research support acknowledged the federal government’s involvement in patents whereas 76% of patents from US government research support only did so.

Third, I investigated if commercial potential of research outcomes was associated with the likelihood of acknowledging government’s research support in resulting patents by utilizing OrangeBook data provided by the US Federal Drug and Food Administration (FDA). I considered patents on FDA-approved drugs recorded in the OrangeBook as those on research outcomes exhibiting significant commercial potential. My analysis confirmed that patents on FDA-approved drugs showed 33% points less likely to acknowledge the government’s research support in patents than their comparison group.

The findings imply that the private stake in the use of the research outcomes is negatively associated with the likelihood of acknowledging government’s research support in patents. Considering that acknowledgment of government’s research support in patents is a legal device to secure governmental right to use research outcomes to serve taxpayer’s interest, the under-credited government’s involvement in patents may limit the potential social benefit from governmental use of research outcomes. The prevalence of underappreciated governments’ research support in patents indicates that current institutions are insufficient to secure government’s discretion to exercise their patent right, which emphasizes that policymakers and scholars need to discuss the way of improving relevant institutions.

ABSTRACT. Research question

What similarities, differences, approaches, and particularities characterize patenting policies and decisions for computer-implemented inventions in the European Union and the Andean Community?

Software is protected by copyright,however, over the past few decades, technological advancements have associated software with new functionalities and creations that open the possibility of claiming protection through the patent regime, due to the fulfillment of the requirements of novelty, inventive level, or industrial application. This is a golden age of technology marked by the great abundance of data and computing capacities, giving rise to technological developments applied to countless areas of science in general and thus to the patenting of what has been called ‘computer-implemented inventions’ (hereinafter CII).

From the legal point of view, TRIPS (WTO. 1994) is one of the most significant legal instruments of the 20th century concerningIP and the multilateral trade system, since it is the affirmation and recognition that nearly all consumer products undergo transformative processes involving raw materials or inputs, culminating in a final product that involves IP rights

However, there are many debates that can arise around the technological change involving CIIs and their legal aspects, as TRIPS does not eliminate legal concerns about the difficulties that many patent offices and courts have in distinguishing patentable from non-patentable subject matter, an issue that has led to the creation of guides to facilitate this work.

On the other hand, middle- and low-income countries consider that protecting software through the patent system entails lengthy prosecution processes that sometimes do not coincide with its life cycle. Additionally, it inhibits competition by generating access barriers, since it hinders interoperability between programs, systems and networks. On the other hand, the high costs involved in patent protection which negatively affect medium-sized and small companies or independent developers when they wish to protect their inventions in different countries (AIPPI, 2017; Saiz, 2022).

Thus the intention of this summary is to present the results of project research in which the scope and limits of Cs were analyzed based on: (a) reviewing its conceptual evolution in the scientific literature indexed in the Scopus database between the years 2015 to 2022; (b) reviewing its legal treatment in regulatory frameworks of multilateral organizations such as the World Trade Organization (WTO) and the World Intellectual Property Organization, as well as the development of policies, standards and guidelines of regional bodies -in the case of the European Union and the Andean Community-; and (c) reviewing some contributions of actors interested in the patent system and relevant to achieve a better understanding of the limits and scope of CIIs.

Methodology

Analyzing and interpreting complex scenarios necessitates examining CIIs through multiple approaches or dimensions, in order to provide greater depth to the reflection on their scope (Phillips and Ritala, 2019). Thus, in this article, the following reflections are put forward: (a) from the conceptual dimension, the literature review enables the analysis of the elements defining computer-implemented invention and distinguishes them from those pertaining to software; (b) from the regulatory dimension, the analysis of international treaties, policies, guidelines and regional mandates, facilitates the identification of characteristics unique to functionalities of computer-implemented inventions protected under patent system and the different levels of development oriented to their protection; c) from the structural dimension, with respect to the actors, a presentation is made of those interested in the development of computer-implemented inventions, in order to understand the dynamics, activities and contributions.

The research was developed from a qualitative approach and an exploratory scope. To achieve the purpose, a literature review was conducted based on the following phases: 1) data collection, 2) data analysis and 3) review report.

Subsequently, in the selection process of documents collected in the Scopus database, the following inclusion criteria were applied: 1) that they were publications between 2015 and 2022; 2) that they included one or more of the admitted keywords; 3) selection by abstract content; and 4) that they were articles published in open access, although those obtained in the databases licensed by the institutions involved were added.

Results

Regarding the conceptual dimension, the evidence suggests that clarity has emerged over time if we compare the definitions of CIIs proposed between 2010 to 2022 in the literature as well as by patent offices.

Regarding the regulatory dimension, some of the key challenges identified, which motivated this research, focused precisely on this component, since the unifying role of multinational organizations on the term CIIs is not evident in the regulatory framework. In other words, the treaties administered by the WIPO and the WTO are not a clear reference on their protection and, instead, jurisprudence is the source of law from which clarity, limits and scope have been granted to what is understood by this type of inventions. This shows that it is especially for countries with a Latin and positivist tradition that the difficulties are accentuated by the regulatory vacuum that this implies.

Regarding the structural dimension, which refers to the actors, the structure shows a relationship in which the linear and descending institutional hierarchy does not provide clarity in terms of legal certainty. The definition of CIIs' limits originates not from international entities but from regional and national governments at two levels: administrative in the patent offices and judicial in tribunals and high courts.

ABSTRACT. Utility model patents and similar second-tier patent rights, referred to as “utility models” for short, are an important yet understudied class of intellectual property rights (IPR) assets (Heikkila, 2023; Kim, Lee, Park & Choo, 2012; Lee, 2024; Prud’homme, 2014). These rights can offer quicker protection than conventional patents, thereby enabling better appropriability of shorter lifecycle innovations (e.g., Heikkila & Lorenz, 2017); protect incremental inventions incapable of meeting the higher standards for conventional patents (e.g., Heikkila & Verba, 2018; Suthersanen, 2006, 2019); and more readily facilitate learning than conventional patents, especially among latecomers seeking to catch up, about how and why to use IPR (Lee & Kim, 2010; Kim et al., 2012). The main mechanism behind these benefits is the greater “accessibility” of utility models compared to conventional patents – whereby the accessibility/inaccessibility of an IP right is determined by the formal institutional rules determining how difficult and costly it is to obtain and maintain from the government (e.g., de Saint-Georges & van Pottelsberghe de la Potterie, 2013; Lemley, 2001; Prud’homme, 2017a; van Pottelsberghe de la Potterie, 2011). Utility models are comparatively more accessible than conventional patents because, on average, they have lower legal granting requirements (e.g., Prud’homme, 2014; Radauer et al., 2015, 2019; Suthersanen, Dutfield, & Chow, 2007).

However, despite these purported benefits, there is scant empirical research examining the impacts of heterogeneities in the accessibility of utility models over time and around the world (Prud’homme, 2017b; Heikkila, 2023). These impacts may include, for example, differences in firms’ decisions to apply for utility models depending on the accessibility/inaccessibility of the regime governing those rights and how it complements or substitutes conventional patent regimes across time and around the world. Examining these issues is important. It can inform lawmakers, in both developing and developed economies (i.e., countries and regions), as well as academics about how to optimize utility model accessibility/inaccessibility – which should enable wider adoption of, and therefore more organizations to benefit from, the rights.

The purpose of the paper summarized in this extended abstract is to evaluate the impacts of heterogeneities in the accessibility of utility models over time and in economies around the world. We do this by, first, conceptualizing the idiosyncratic core components of utility model accessibility/inaccessibility: the height of the requirements governing the rights, the strictness the requirements’ administration, and associated fees. Then, we hypothesize and test how differences in the accessibility/inaccessibility of utility model regimes impacts filings of utility models and conventional patents. Our research is based on a quantitative UM Inaccessibility Index proxying the laws governing utility model regimes in 25 economies over three and a half decades, its interactions with conventional patent strength, and resulting impacts on utility model filings and conventional patent filings.

Our research offers two main contributions. First, we complement the sparse yet highly valuable empirical literature indicating that there may be a negative effect of more inaccessible IPR regimes on usage of those regimes (de Saint-Georges & van Pottelsberghe de la Potterie, 2013). We do so by offering evidence of an inverted U-curve between utility model regime inaccessibility and utility model usage: a moderate amount of inaccessibility may lead to more filings than a lot or a little inaccessibility – at least over time and across multiple economies, ceteris paribus. This indicates that prior hypothesizing about the impact of less accessible conventional patent regimes (de Saint-Georges & van Pottelsberghe de la Potterie, 2013) is not just valid in that context but also in the context of utility model regimes. This is especially important because there are much greater heterogeneities in the accessibility of utility model regimes than conventional patent regimes around the world, given that the Agreement on Trade-Related Aspects of Intellectual Property (TRIPs) sets minimum standards for the latter rights but not for the former (e.g., Janis, 1999; Gross Russe-Khan, 2012). Moreover, our work contributes helpful nuance to de Saint-Georges & van Pottelsberghe de la Potterie (2013) by showing that the relationship between IPR regime inaccessibility and IPR filing might be curvilinear (rather than linear), and that it persists over time (rather than just across a cross-section of time).

Second, we complement the sparse yet highly valuable empirical literature on the substitutive relationship between utility models and conventional patents over time (Lee, 2024; Lee & Kim, 2010; Heikkila, 2017; Kim et al., 2012). Prior literature has not yet had the chance to examine the role that changing utility model regime inaccessibility may play in substituting conventional patents for utility models across economies and over time. Nor has it had the chance to empirically examine the interaction of such inaccessibility with conventional patent regime strength over time and across economies. By analyzing what appears to be, to the best of our knowledge, the first quantitative indexes of utility model regime inaccessibility around the world and over time (the indicators for which are presented in Table 1), and examining how they interact with conventional patent regime strength (Park, 2008), we therefore add important nuance to this line of research. We show that firms may substitute other means of appropriability for utility models, such as conventional patents, when a utility model regime is more inaccessible and when conventional patent regimes offer stronger rights.

Table 1: Indicators for UM Inaccessibility Index Indicator Sub-indicators Sub-indicator scoring Sub-indicator weighting Highest score possible 1. Novelty 1. Absolute novelty 3 2 6 2. Relative novelty 2 2 3. Local novelty 1 2 4. None 0 2 2. Inventive step Requirement of any kind for utility models or not 3 or 0 1.75 5.25 3. Examination 1. Patentability of subject matter examined in any way or not 3 or 1 0.1 3 2. Industrial applicability examined in any way or not 3 or 0 0.15 3. Novelty examined in any way or not 3 0.3 4. Inventive step examined in any way or not 3 0.45 4. Opposition and related mechanisms 1. Pre-grant mechanism of any kind or not 3 or 0 0.4 3 2. Post-grant mechanism of any kind or not 3 or 0 0.6

ABSTRACT. Since Price presented his postulates on the shift to Big Science (1963), many have attempted at describing the social dynamics surrounding the development of scientific breakthroughs and the implementation of policies directed at fostering a productive research environment. Decades of work on interdisciplinary knowledge creation has advanced our understanding on knowledge-related dynamics across disciplines (Gibbons et al., 1994). Their challenges have been well documented in scientific collaboration studies (Bolduc, Knox & Ristroph, 2023; Klein, 2014). These address the dynamics of the inter-discipline, pointing to differences in language, research and team norms, and other disciplinary characteristics for working across boundaries (Leahey, Beckman & Stanko, 2017). These disciplinary distinctions are acknowledged, but not systematically examined.

From a theoretical perspective, Richard Whitley (2000) built a framework describing social and intellectual organization across fields based on two key concepts: mutual dependence and task uncertainty. While the former refers to the need for external validation, the latter focuses on labor distribution. Whitley suggests that different fields will have varying degrees of dependence and task uncertainty, proposing seven typologies of working structure. The dynamics of the research process are affected by myriad factors. Whether the research is more computational in nature, lab or field-research based, and reliant on specialized equipment or facilities are factors that drive the organization of scientific activities. Where a scientific team is based and where the scientists work introduce institutional, regulatory, and other factors that shape the nature of work.

Empirical studies on the social composition and dynamics of research activities have primarily focused on issues such as team size, power dynamics, collaboration or team composition and structure. This is due to the difficulty to examine the internal dynamics of team science at a large scale. These studies tend to rely on bibliographic data, applying bibliometric indicators based on author position, and combining it with individual traits such as gender or past trajectory. Science teams challenge current notions of credit in the scientific ecosystem (Walsh & Lee, 2015), as it becomes increasingly difficult to distribute it based solely on authorship position (Biagioli & Galison, 2013). In the last decade, the expansion of contribution statements across publication records allows for the first time to explore the relation between author order, time size and distribution of tasks, based on the self-reported statements of authors themselves. Benefiting from these novel data, we ask: how does intellectual leadership and research task distribution vary across disciplinary areas?

Data and Methods We analyze a set of almost 700,000 publications using the Contributor Role Taxonomy (CRediT) based on a set of Elsevier and PLOS journals to study the organization of the sciences across fields (González-Salmón et al., 2024). CRediT is increasingly being used by journals to catalog author contributions. Co-authors indicate their specific contribution out of 14 possible roles.

Building upon Whitley’s framework, we map the levels of task uncertainty and mutual dependence across and within fields. Considering that a paper reflects the outcome of a collective endeavor, we distinguish between scientific collaborations and highly stratified and organized team science with explicit structures. We apply agglomerative hierarchical clustering to publications by field and characterize the resulting clusters based on the team size and two measures of IDR, based on references and based on authors (Pinheiro et al., 2021).

Preliminary and Expected Findings So far we have analyzed the field of Medicine, but will extend our analysis to all fields prior to the conference. After applying bootstrapping and using different resolutions, our initial results suggest that there are four collaboration structures. These show different organizational structures related to differing modes of research. However, in terms of IDR, we find no meaningful differences across groups.

Cluster 1. All team members contribute to most tasks, with a moderate contribution on administration (e.g., funding). Papers are authored by a low number of researchers (avg. team size: 4.7). Cluster 2. Teams are characterized by a high distribution on technical tasks with a shared contribution on conceptual tasks. (avg. team size: 5.9). Cluster 3. Teams are larger with a higher distribution of labor for all tasks. (avg. team size: 7.5) Cluster 4. High distribution of tasks, although most members participate on the Review & Edit contribution (avg. team size: 8.2).

We will expand our analysis to all fields of research, exploring contribution co-occurrence by fields, and adding further variables such as seniority, gender, collaboration type and funding (based on funding acknowledgements).

Next steps The results of our research contribute to the understanding of collaboration and team dynamics in publishing activities. Because CRedit asks co-authors to indicate their role in the initiation, conduct, and communication of the research, it enables us to examine these activities across disciplines. These data greatly extend what has been possible to surmise from author order, but have limitations (Sauermann & Haeussler, 2017). There may be different approaches to how author contributions are reportedb However, given the size of our dataset, we can discern organizational structures across and within disciplines to understand the internal dynamics of team science.

Bibliography Bolduc, S., Knox, J., & Ristroph, E. B. (2022). Evaluating team dynamics in interdisciplinary science teams. Higher Education Evaluation and Development, 17(2), 70-81. https://doi.org/10.1108/HEED-10-2021-0069 Gibbons, M., Limoges, C., Nowotny, H., Schwartzman, S., Scott, P., & Trow, M. (1994). The New Production of Knowledge: The Dynamics of Science and Research in Contemporary Societies. SAGE. González-Salmón, E., Di Césare, V., Xiao, A., & Robinson-Garcia, N. (2024). Beyond authorship: Analyzing disciplinary patterns of contribution statements using the CRediT taxonomy. Zenodo. https://doi.org/10.5281/zenodo.14168888 Klein, J. T. (2005). Interdisciplinary Teamwork: The Dynamics of Collaboration and Integration. In Interdisciplinary Collaboration. Psychology Press, p. 23-50. Leahey, E., Beckman, C. M., & Stanko, T. L. (2017). Prominent but Less Productive: The Impact of Interdisciplinarity on Scientists’ Research*. Administrative Science Quarterly, 62(1), 105-139. https://doi.org/10.1177/0001839216665364 Price, D. J. de S. (1963). Little science, big science. Columbia University Press New York. http://www.garfield.library.upenn.edu/lilscibi.html Sauermann, H., & Haeussler, C. (2017). Authorship and contribution disclosures. Science Advances, 3(11), e1700404. https://doi.org/10.1126/sciadv.1700404 Whitley, R. (2000). The intellectual and social organization of the sciences (2.a ed.). Oxford University Press.

ABSTRACT. INTRODUCTION I examine the effect of interdisciplinarity on attention to research by policy makers. Publications categorized by multiple fields of research have much higher rates of citation in policy documents. This study considers the policymaking landscape in the US and EU and the publications cited in governmental policy documents, comparing alpha-beta collaborative research with single-field research. In addition to this analysis, in order to determine whether funding mechanisms play a role in encouraging policy-relevant interdisciplinary research, two specific initiatives (the ERC Synergy Grants and the NSF INSPIRE mechanism) are evaluated, comparing the policy uptake of publications arising from research funded by these grants to that of other similarly-sized, single-field research projects supported by the ERC and NSF. Patterns over time and differences between fields of research are each considered.

BACKGROUND In previous work (Campaign for Social Science 2023), we have found that the uptake of research by UK governmental policymaking organizations was proportionally higher for those publications which cross the STEM/Social Sciences divide (alpha-beta research). Specht & Crowston (2022) looked at interdisciplinarity and considered the role of multiple qualitative favors, as well as assessing impact by comparing the quantity of research output and the sum of scientific citations. I have drawn on Taylor (2023) for overall baselines for normal citation timelines for research output in policy documents.

DATA This analysis is produced using the Dimensions dataset, which consists of tables of metadata about (among other data types) nearly 150 million publications, over 7 million grants, and 2.5 million policy documents. The relationships between these tables are provided via linking fields (supporting_grant_ids, (citing) publications, etc.). The Dimensions dataset is further enriched with machine learning classification, organization metadata, and funder-provided grant details, including the fields of research of publications, the type of policymaking organization, and the grant program for each awarded grant. The connection between 7.4 million grants and their over 12 million resulting publications have been established by looking at the grants mentioned in the text of the publications. Citations of publications in policy documents are identified by using persistent identifiers (DOIs, PubMedIDs, etc.) and links (handles and URLs) to publications in the policy document text. The Dimensions dataset has the advantage of being freely available for scientometric research projects, which means that this study can be reproduced and interrogated, and future patterns of policy document citation can be investigated in comparison to this work.

METHODS The comprehensive nature of the dataset and the capabilities of Google's BigQuery make it possible to do a descriptive analysis, making sampling of the dataset unnecessary. The study population includes, in the first case, global research output categorized by number of associated fields of research.

In order to establish whether the interdisciplinarity is a result of collaboration or design, a disciplinary profile is calculated for the co-authors of these publications: primarily STEM, primarily social science, or primarily interdisciplinary. Trends observed in the temporal analysis are compared across research with various collaboration patterns.

The distribution of policy attention to research over time (from time of publication to time of policy attention) shows a peak in first mentions of research by policy documents in the second year after publication, with the number of new first mentions dropping below the frequency of the first year only after year 7. This makes estimating new trends in policymaking a long-term affair. The first temporal analysis therefore considers the impact over time of research published between 2015-2017, and the second compares the results of the first two years of policy citations of that research with the first two years of policy citations for research published in 2022.

DELINEATION OF THE STUDY POPULATION The first corpus of publications is the 114 million publications categorized by field of research in the Dimensions dataset. The rates of policy citation for these, by number of fields and by interdisciplinarity of authors, are compared over time. The second corpus of publications includes those published between 2015-2017 and categorized by field of research. This consists of just over 11 million publications. The third corpus of publications consists of documents published in 2022 and categorized by field of research, of which there are 5,686,668 publications. Finally, I consider publications arising from the NSF- and ERC-supported interdisciplinary grants and their resulting publications. Of 193 grants recorded as belonging to the INSPIRE or Synergy programs in Dimensions, 173 have associated resulting publications. Of those, 127 publications resulting from 27 grants have been cited in 366 policy documents.

CONSEQUENCES FOR RESEARCH AND FUNDING POLICY The results of this analysis have shown that policy documents cite interdisciplinary research at a much higher rate. The findings provide evidence for the importance of interdisciplinary co-operation from the point of research design through co-authorship. This research suggests that policies that encourage cross-disciplinary collaboration have a positive influence on the production of policy-relevant research. By providing data to support the effect of such policies, this research should empower funding agencies to develop more impactful mechanisms to support policy-relevant research.

CODE AVAILABILITY AND DATA ACCESS The code used for this study will be made publicly available on Figshare. All of the data used are accessible with a Dimensions license or scientometric access to Dimensions.

REFERENCES Campaign for Social Science Reimagining the Recipe for Research & Innovation: The Secret Sauce of Social Science. First Edition. SAGE Publications Ltd, 2024. https://doi.org/10.4135/9781529681345.

Specht, A. & Crowston, K. (2022) Interdisciplinary collaboration from diverse science teams can produce significant outcomes. PLoS ONE 17(11): e0278043. https://doi.org/10.1371/journal.pone.0278043

Taylor, M. (2023) Slow, slow, quick, quick, slow: five altmetric sources observed over a decade show evolving trends, by research age, attention source maturity and open access status. Scientometrics 128, 2175–2200. https://doi.org/10.1007/s11192-023-04653-1

ABSTRACT. Interdisciplinarity fosters innovation by bridging diverse methodologies and knowledge domains. The increasing complexity of global challenges has amplified the role of interdisciplinarity in advancing scientific innovation. There are also widespread efforts to promote interdisciplinarity in U.S. universities and PhD programs. While interdisciplinary PhDs are shown to have long-term gains in collaboration and productivity, they may face obstacles during early career transitions, such as reduced career opportunities. This study examines the impact of PhD interdisciplinarity on academic placement and long-term career trajectories, providing insights into how interdisciplinary training interacts with institutional hierarchies and hiring practices. To explore this issue, we used data sourced from the Academic Analytics Research Center (AARC), encompassing tenure-track faculty records from U.S. PhD-granting institutions between 2011 and 2020. PhD information, including department, dissertation titles, and advisor details, was supplemented from the ProQuest Dissertation database. To enhance coverage, machine learning models were employed to predict missing PhD subfields using dissertation titles. Publication records were further enriched by matching DOIs with Clarivate’s Web of Science, enabling analyses of interdisciplinarity, productivity, and citation patterns. The final dataset comprised 32,977 faculty members, 1,038,518 publications, and detailed advisor and collaborator networks. Each faculty member’s interdisciplinarity during their PhD (PhD interdisciplinarity) is the median interdisciplinarity score of their PhD-stage publications. We measured paper-level interdisciplinarity score using the Rao-Stirling index, which quantifies the diversity of references in academic publications. This index considers both the disciplinary variety of references and their similarities. References were categorized into 144 disciplines using the Web of Science classification. University rankings were based on a "production rank", which evaluates an institution's ability to place graduates into higher-ranked faculty positions within their subfields. Institutions were ranked by percentile, with the top 10% identified as those in the 90th percentile or higher within each subfield. By descriptive statistics, we found that faculty members with high PhD interdisciplinarity face disadvantages in securing faculty positions at top-ranked institutions. This trend is evident across most fields except Math & Computing. Faculty at top institutions are more likely to have lower interdisciplinarity, suggesting a preference for candidates with a narrower disciplinary focus. Analysis of hiring trends reveals that interdisciplinarity remains stable across lower-ranked institutions but declines sharply for placements at top-ranked institutions. Logistic regression models controlling for factors like PhD institution rank, graduation year, individual features, advisor features, and collaborator features confirm that higher interdisciplinarity correlates with lower odds of placement at top-ranked institutions, especially within fields like Physics & Engineering, Life & Earth Sciences, and Social Sciences. Quantile regression further demonstrates that this disadvantage is concentrated in higher institutional ranks, highlighting structural barriers for interdisciplinary scholars in the most competitive academic environments. Moreover, we found such a challenge for interdisciplinary PhDs happens if they seek job in their original fields. Top institutions tend to hire a higher proportion of cross-field faculty—those transitioning into new or distant fields—than lower-ranked institutions, reflecting efforts to diversify academic expertise. However, interdisciplinary PhDs face reduced chances of being hired for faculty roles within their original discipline at these top institutions. Multinomial logistic regression shows that as interdisciplinarity increases, the likelihood of securing same-field placements at top institutions decreases significantly. In contrast, cross-field placements at top institutions are less affected or may even increase with interdisciplinarity in fields like Life & Earth Sciences and Social Sciences. This suggests that top institutions value interdisciplinarity more for cross-field roles than for maintaining disciplinary continuity, creating barriers for interdisciplinary scholars who remain in their primary field. By measuring the research agenda alignment between our sample and each institution’s existing faculty, we see faculty members with low Ph.D. interdisciplinarity generally have smaller deviations from the research of existing faculty, and this alignment is more pronounced at higher-ranked institutions. In contrast, faculty members with high Ph.D. interdisciplinarity show consistently higher deviations. This misalignment shows the perceived lack of "academic fit" between highly interdisciplinary research and the traditional focus areas of top-ranked institutions, further reinforcing systemic barriers for these scholars. What are the potential consequences of such challenges for interdisciplinary PhDs? We first examined it from a gender perspective. We notice that women PhDs exhibit higher levels of interdisciplinarity than their male counterparts across most fields. However, this heightened interdisciplinarity correlates with reduced opportunities at top-ranked institutions, potentially exacerbating existing gender disparities. Propensity score matching indicates that accounting for interdisciplinarity mitigates these disparities, even creating a slight advantage for women in some fields. This indicates that the structural challenges interdisciplinary PhDs face in faculty hiring may contribute to the existing gender disparities in academia. Furthermore, using event study analysis, we identified that interdisciplinary faculty have higher publication rates over the decade following PhD graduation, particularly in Life & Earth, Biology & Health, and Social Sciences. Comparing the top-ranked and other institutions, we see this productivity advantage is even more pronounced at top-ranked institutions. Therefore, investing in and supporting interdisciplinary scholars at top institutions may yield long-term gains in research output, with the current hiring barriers potentially leading to a loss in knowledge production. Overall, our findings highlight a paradox: while interdisciplinarity is heralded as vital for addressing multifaceted societal challenges, systemic biases in faculty hiring constrain the career advancement of interdisciplinary scholars. Institutional hiring practices favor narrowly defined disciplinary expertise, potentially stifling innovation and perpetuating inequities. The gendered dimensions of interdisciplinarity further prove the need for targeted interventions to support diverse scholars. To fully leverage the potential of interdisciplinary research, academic institutions need to adapt hiring practices to value interdisciplinary expertise. Initiatives such as interdisciplinary clusters, dual appointments, and funding programs can create pathways for interdisciplinary scholars to thrive. By fostering a more inclusive and strategically aligned approach to interdisciplinarity, academic institutions can better address complex, cross-disciplinary challenges and ensure the equitable advancement of diverse scholars.

ABSTRACT. Academic activity has increasingly become characterized by collective and collaborative work. This process of the collectivization of science (Ben-David, 1984)—which is not new—can be understood as the result of the intrinsic dynamics of science (its advancement and growing complexity) on the one hand, and on the other, as influenced by research policies. These policies foster strategic groupings to enhance opportunities for funding linked to policy priorities, which demand complementarity and interactions between knowledge producers and users (Bianco and Sutz, 2005).

Empirical studies on research groups can be classified into three main approaches: 1.Institutional or national surveys, with notable cases including: Brazil, where the National Council for Scientific and Technological Development (CNPq) has conducted a national census of research groups since 1992. These groups are defined as hierarchical collectives organized around common research lines, emphasizing scientific leadership, shared resources, and continuous output; Colombia, where Colciencias invites groups for evaluation to facilitate funding access. These groups are defined as sets of individuals strategically working on common problems, producing tangible and verifiable results through formalized projects. 2.Studies on internal dynamics, focused on analyzing the specific processes and practices of groups in certain fields, such as chemistry in U.S. universities or molecular biology in Europe. These groups are identified as entities recognized by peers, collaborators, or funders. 3.Exploratory diagnostics, examining the state of research within specific institutions. Examples include studies at the University of Barcelona on scientific evolution (1985–1990) and at the University of Buenos Aires on academic output, human resource training, and organizational aspects of 200 research groups.These approaches underscore the importance of groups as knowledge production units, addressing their identification, evaluation, and organizational dynamics. (Bianco and Sutz, 2005).

More recently, Huang et al. (2023) emphasize that collaborative teamwork for addressing complex scientific and social issues has become increasingly common. This demands the integration of knowledge and the complementarity of skills and resources across cultural and organizational boundaries. According to these authors, a team is defined as a formal group of individuals who interact interdependently to achieve common goals. They highlight three main types of teams: 1. Traditional teams, based on existing administrative structures; 2. Virtual teams, formed through collaborative relationships that materialize in co-authorships; 3. Temporary teams, created for specific projects that dissolve upon completion. Teams can further be classified as unidisciplinary, multidisciplinary, interdisciplinary, or transdisciplinary.

Team science, an interdisciplinary and emerging field of study, has drawn academic interest in analyzing strategies to strengthen research teams. It examines how scientific teams, research centers, and institutes of various sizes and structures are formed. Researchers explore team formation and performance from different perspectives, such as: Considering co-authors of an article as a team; Defining project members as a research team; Viewing university centers as research groups. These analyses focus on aspects like knowledge contribution, member interests, performance, collaboration, and knowledge management. (Cheng, Zou, & Zheng, 2024). These authors, aiming to improve policies and practices to strengthen collaborative research in academia, focus on the identification and classification of research teams in the field of materials science at two universities in Asia to address the diversity in their structure, dynamics, and collaboration. They define teams as groups of individuals collaborating to achieve common goals, sharing resources and information. According to their findings, there are different types of teams based on their scale, characteristics, and contexts; these teams are often interdisciplinary and reflect the diversity of projects and collaborations within universities. The "key members"—defined by their high productivity and centrality in co-authorships—play a crucial role in the structure and performance of the teams.

However, despite these efforts, academic reflection on research groups and their internal dynamics remains limited, especially regarding the development of methodologies for their systematic identification across various fields of knowledge (Cheng, Zou, & Zheng, 2024).

In the context of this work, the University of the Republic (Udelar) in Uruguay first proposed fostering collective research work in 1994. This led to questioning what constitutes a research group. In 2001, Udelar conducted its first self-identification survey for research groups to reformulate policies promoted by the Sectoral Commission for Scientific Research (CSIC) (Bianco y Sutz, 2005). This effort aimed to integrate research groups into these policies by defining them as collectives of two or more individuals working jointly on shared research topics (Unidad Académica-CSIC, 2003).

In 2010, Udelar launched a new self-identification survey (Ardanche, Bianco y Tomassini, 2014) and created the CSIC’s R&D Groups Program, which has since held regular calls every four years (2010, 2014, 2018, and 2022). This program offers competitive funding for research groups to strengthen their agendas, teaching activities, and researcher training under better conditions.

This study presents the main findings from over a decade of implementing this policy instrument aimed at strengthening research groups in all areas of knowledge cultivated at Udelar. The study focuses on the problematisation of the definition of what a research group is, what different types of groups there are, their attributes, and examining their evolution over time. It is relevant to answer these questions with new knowledge in order to delimit the target population of such an instrument, as well as to assess the need to expand its modalities to respond to possible differences between groups (e.g. according to their degree of consolidation, size, age, access to previous funding, etc.). It employs quantitative and qualitative methods, including analyzing databases of 1,070 self-identified groups and 158 funded groups (2010–2024) and conducting document reviews and focus groups with evaluators and group members. The databases include information on: lines of research, disciplines, motives and year of creation of the groups, disciplines, their activities, international and national academic colaborations and links with other extra-academic actors, projects, funding, production, characteristics of its members (sex, training, academic trajectory, etc.) among others.

It seeks to draw lessons for continuous policy improvement based on evidence and systematic reflection of its results. In this way, this paper seeks to contribute to knowledge at the intersection of studies on the dynamics of science itself - and its modes of production - and research policy.

ABSTRACT. OpenAI's demonstration of multimodal artificial intelligence (AI) services, featuring a blind person utilizing a mobile application for taxi hailing, illustrates the technology's capacity to simulate human-like assistance through contextual intelligence. Nevertheless, the development and deployment of such capabilities are contingent upon significant computational resources, including extensive datasets and advanced computing resources like GPUs, which may amplify existing socio-economic inequalities. This talk examines the intricate relationship between the rapid advancement of AI and social inclusion, with a particular focus on marginalized communities, such as older adults and persons with disabilities. I will provide the implications of concentrated investment and accelerated development within the AI sector, delineating both the potential for enhanced accessibility and the risk of heightened digital exclusion. Specifically, this talk will scrutinize recent legislative initiatives designed to promote inclusive AI, with a primary emphasis on the Digital Inclusion Act, passed by the Korean National Assembly in December 2024. Furthermore, I will explore the Act's legislative process, including its multi-stakeholder consultations and background, as well as the unfolding details of its subordinate legislation and implementation strategies. Finally, I will address the ongoing ethical discourse surrounding the integration of AI within social infrastructure, considering the long-term sustainability of inclusive AI initiatives and the imperative for continuous evaluation of their societal impact.

ABSTRACT. Artificial intelligence (AI) holds significant promise in enhancing the lives of neurodiverse individuals—those with neurological variations such as autism, ADHD, and dyslexia—by providing personalized support in communication, daily routines, and social participation. Recent studies demonstrate how AI-driven services can empower these users; for instance, large language model-based conversational agents offer autistic individuals a non-judgmental source of advice and coaching for everyday tasks, fostering greater independence (Cha et al. 2021, Choi et al. 2024). Similarly, AI-powered tools designed explicitly to accommodate neurodiverse users, such as gamified scheduling apps and adaptive social interaction platforms, have shown improvements in user engagement, task management, and self-efficacy (Kim et al., 2023, Kim et al., 2024). These findings underscore critical ethical considerations in AI development for marginalized communities. Embracing the social model of disability—which frames disability as arising from societal barriers rather than inherent deficits—requires designers to recognize cognitive variations as natural diversity to support rather than deficits to correct. For example, autistic individuals navigating online dating seek clarity on social conventions yet prefer preserving their unique characteristics, even if these diverge from neurotypical norms (Choi et al., 2023). However, mainstream matchmaking algorithms aligned with majority preferences inadvertently exclude such users, prompting implicit compromises in their authentic self-expression. Similarly, neurodivergent YouTube creators strategically portray their disability identities to educate the public but frequently find algorithms insufficiently supportive, diminishing their visibility and outreach (Choi et al., 2022). Considering that current AI systems often fail to adequately reflect neurodivergent users' self-defined priorities and inadvertently marginalize their authentic voices and perspectives, it is essential to explore new ethical design approaches—actively integrating neurodiverse individuals' unique needs, values, and identities, while consciously amplifying rather than suppressing their contributions within algorithmic decision-making processes. To foster responsible and inclusive AI innovation, this research advocates for a paradigm shift in the development of AI systems for neurodiverse populations. Co-design and participatory methodologies involving neurodivergent individuals as a key informant are essential for ensuring AI systems authentically represent their lived experiences, needs, and values. Transparency and fairness in algorithmic decision-making processes also must be critically examined to prevent systemic disadvantage or stigmatization of neurodiverse users. Moving forward, comprehensive frameworks and AI literacy initiatives should be established to: (1) recognize and respect neurodiversity as a cultural identity with unique experiences and perspectives, rather than a deficit or disability; (2) ensure AI systems preserve users’ distinct identities, values, and agency by not implicitly enforcing neurotypical standards; and (3) facilitate reciprocal understanding and bridge societal divides by enabling neurotypical populations to appreciate neurodiverse strengths and characteristics (Kim et al., 2024). This approach aligns with responsible AI principles—prioritizing equity, autonomy, and cultural sensitivity—and charts a path for AI innovation that uplifts marginalized communities. References Kim, S. I., Jang, S. Y., Kim, T., Kim, B., Jeong, D., Noh, T., ... & Kim, J. G. (2024). Promoting self-efficacy of individuals with autism in practicing social skills in the workplace using virtual reality and physiological sensors: Mixed methods study. JMIR Formative Research, 8, e52157. Choi, D., Lee, S., Kim, S. I., Lee, K., Yoo, H. J., Lee, S., & Hong, H. (2024, May). Unlock life with a Chat (GPT): Integrating conversational AI with large language models into everyday lives of autistic individuals. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (pp. 1-17). Kim, B., Jeong, D., Hong, H., & Han, K. (2024, May). Narrating routines through game dynamics: Impact of a gamified routine management app for autistic individuals. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (pp. 1-15). Kim, B., Kim, S. I., Park, S., Yoo, H. J., Hong, H., & Han, K. (2023, April). RoutineAid: externalizing key design elements to support daily routines of individuals with autism. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (pp. 1-18). Choi, D., Kim, S. I., Lee, S., Lim, H., Yoo, H. J., & Hong, H. (2023, April). Love on the Spectrum: Toward Inclusive Online Dating Experience of Autistic Individuals. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (pp. 1-15). Cha, I., Kim, S. I., Hong, H., Yoo, H., & Lim, Y. K. (2021, May). Exploring the use of a voice-based conversational agent to empower adolescents with autism spectrum disorder. In Proceedings of the 2021 CHI conference on human factors in computing systems (pp. 1-15).

ABSTRACT. AI offers innovative approaches to transform social service delivery and coordination, with significant potential to address persistent, long-standing inequities. Applications already in development include predictive risk modeling to identify families and children in need (Ahn et al., 2024; Chouldechova et al., 2018) and simulation modeling to optimize resource allocation across service systems (Mashiat et al., 2024). Despite these advances, there remains a critical gap in the development of comprehensive governance frameworks specifically designed for AI deployment in public sector social services.

Private sector AI governance frameworks prioritize organizational value creation, technical performance, and regulatory compliances (Batool et al., 2025; Birkstedt et al., 2023). These general frameworks organize governance around ensuring AI aligns with corporate strategies by emphasizing data management, algorithm functionality, and risk mitigation primarily from business and legal perspectives. In contrast, public sector frameworks shift focus toward procedural fairness, institutional accountability, and collaborative governance for public interest. For instance, Wirtz et al.'s (2020) model extends beyond private sector concerns by explicitly addressing societal impacts, ethical considerations, and multi-stakeholder policy development, acknowledging unique public sector responsibilities while balancing innovation with public protection.

When social services target inequality reduction among marginalized communities, existing public sector AI governance frameworks require further adaptation. These specialized contexts demand incorporation of structural inequities analysis, trauma-informed principles, and social work ethics into governance processes (Ahn et al., 2025). Effective governance for these applications must establish mechanisms that address power differentials and systemic inequalities beyond what generalized public sector models typically provide (Bricout & Goldkind, 2024). Without this critical extension of public sector governance for social service contexts, AI implementation risks reinforcing existing inequities rather than mitigating them. References Ahn, E., An, R., Jonson-Reid, M., & Palmer, L. (2024). Leveraging machine learning for effective child maltreatment prevention: A case study of home visiting service assessments. Child Abuse & Neglect, 151, 106706. https://doi.org/10.1016/j.chiabu.2024.106706 Ahn, E., Choi, M., Fowler, P., & Song, I. (2025). Artificial Intelligence (AI) Literacy for Social Work: Implications for Core Competencies. Journal of the Society for Social Work and Research, 735187. https://doi.org/10.1086/735187 Batool, A., Zowghi, D., & Bano, M. (2025). AI governance: A systematic literature review. AI and Ethics. https://doi.org/10.1007/s43681-024-00653-w Birkstedt, T., Minkkinen, M., Tandon, A., & Mäntymäki, M. (2023). AI governance: Themes, knowledge gaps and future agendas. Internet Research, 33(7), 133–167. https://doi.org/10.1108/INTR-01-2022-0042 Bricout, J., & Goldkind, L. (2024). Meeting the Grand Challenge to Harness Technology for Social Good: Social Workers Support Vulnerable Community Members’ Rights to Access Safe, Beneficent and Effective AI. American Academy of Social Work & Social Welfare. http://grandchallengesforsocialwork.org/ Chouldechova, A., Putnam-Hornstein, E., Benavides-Prado, D., Fialko, O., & Vaithianathan, R. (2018). A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions. Proceedings of Machine Learning Research, 81, 1–15. Mashiat, T., DiChristofano, A., Fowler, P. J., & Das, S. (2024). Beyond Eviction Prediction: Leveraging Local Spatiotemporal Public Records to Inform Action. The 2024 ACM Conference on Fairness, Accountability, and Transparency, 1383–1394. https://doi.org/10.1145/3630106.3658978 Wirtz, B. W., Weyerer, J. C., & Sturm, B. J. (2020). The Dark Sides of Artificial Intelligence: An Integrated AI Governance Framework for Public Administration. International Journal of Public Administration, 43(9), 818–829. https://doi.org/10.1080/01900692.2020.1749851

ABSTRACT. Public-private partnerships play a crucial role in leveraging AI to address social challenges. While private companies possess advanced technological capabilities, public institutions provide the organizational infrastructure and direct engagement with marginalized populations, such as older adults and individuals with disabilities. This talk will present case studies of such partnerships in South Korea, highlighting collaborative efforts between SK Telecom and various non-profit and governmental organizations. SK Telecom, a global leader in AI and a key player in entrepreneurship with a strong focus on Environmental, Social, and Governance (ESG) principles, has pioneered several AI-driven initiatives to support vulnerable communities. One such initiative is AI Call Care, an AI-powered monitoring service designed to enhance the well-being of older adults. The successful implementation of this service has been made possible through long-term collaborations with local non-profit organizations and government agencies. Additionally, SK Telecom recently introduced Carevia, an AI-driven initiative that utilizes vision-based AI to analyze and manage challenging behaviors in individuals with developmental disabilities. Another notable innovation is Sullivan EYE, a multimodal generative AI technology jointly developed with Tuat, a tech startup, which recognizes images, text, and faces to provide verbal descriptions, assisting individuals with visual impairments. By examining these cases, this talk will explore the impact of AI-driven public-private partnerships in fostering social inclusion and improving the quality of life for marginalized populations.

ABSTRACT. One of the most recent challenges posed by the European Union is the transition towards environmental friendly technologies to address major challenges such as the conflicting trends of a growing population and the limited amount of natural resources (European Commission, 2023). In light of this, a strategy is proposed at the European level involving a variety of organizations (e.g. firms, universities and research centers) to renew their knowledge and skill sets to be able to develop these technologies (JRC, 2022). Germany as well as other European countries are currently introducing new green technologies to meet the standards imposed by EU. Given its diverse and broad research infrastructure comprehending both public (universities, universities of applied sciences, technical universities and research institutes) as well as private entities (Graf & Menter, 2022), Germany is a particular interesting research case. Green technologies are complex in their nature and they rely on a higher extent on developments put forward in other technologies (Barbieri et al., 2020, 2023). The literature on technological development focuses mainly on knowledge recombination activities put forward by inventors (Cassi & Plunket, 2015). Mainly due to data limitations, only recently new studies shifted towards assessing the recombination impact of organizations for new technological advancements (e.g. Wanzenböck et al., 2024). We continue this line of research investigating the, potentially different, roles of public and private organizations in making new knowledge combinations with a particular focus on green technologies given their strategic importance. In particular, we focus on a combinatorial function called uniqueness regarded as important for knowledge recombination in general (Su et al., 2013) and even more in the case of green technologies (Barbieri et al., 2020). After identifying from which typology of organization the majority of unique combinations in the green realm come from, we also analyze their impact on subsequent inventions. We further distinguish between impact on subsequent green technologies or on other non-green technologies. Particularly in the context of a green transition, it is important to assess if the new knowledge produced by those inventions is also subsequently re-utilized or not, and in which knowledge fields.

In general, the development of new technologies passes through a process called technological recombination. Knowledge elements (both old and new) are combined in new ways to produce technological advancements (Weitzman, 1998). Different bits of knowledge necessary to produce new technologies employed in a region are called knowledge base (Kalthaus, 2020). The inventors are the ultimate agents which create these new combinations. However, the unit which provides the resources (both tangible and intangible) and ultimately profit (in terms of reputation or revenues) from those inventions are the organizations such as firms, universities or research institutes. But, different types of organizations have different propensities to combine pieces of knowledge due to their orientation towards basic or applied research. Basic research, in the classic view, is focused on introducing new revolutionary technological combinations without a clear application. Whereas applied research is usually redefining knowledge that is already existing to produce and commercialize products and services. Moreover, basic research is usually regarded as more difficult to be performed but with a higher impact on subsequent inventions. Whereas, applied research is usually easier to perform but with a limited impact on further knowledge development (Graf & Menter, 2022). Thus, given those properties we expect to observe differences in recombination activities of private and public organizations with regard to green technologies. In fact, by relying to a larger extent on developments put forward in other technological realms (Barbieri et al., 2020) we expect also to find different effects on subsequent inventions when considering different organizations.

We use OECD REGPAT September 2024 for identifying inventive activities. The focus is on the patents which have at least one inventor with a German address. The considered time period is between 1990 and 2021. We use CPC classes at 5 digit level to identify the knowledge elements. Following Graf & Menter (2022) using the Regpat HAN database, patents are associated using an algorithm to the organizations that filed them. Green technologies are captured by all patents with at least one class falling in the three-digit CPC classification Y02. To measure the uniqueness of combination, we rely on a very well known indicator in Social Network Analysis called Redundancy Coefficient. This indicator is able to capture the extent to which the technological combinations created by a patent are unique in the entire German knowledge base. Finally, in line with other literature in the field to assess the impact on subsequent inventions we rely on forward citations.

In general, we contribute to the literature on technology recombination and more specifically on the literature about green technology development. Our results are also important for policymakers, who should consider how organizations can renew and adapt their knowledge bases in order to accelerate the transition to green technologies.

References Barbieri, N., Marzucchi, A., & Rizzo, U. (2020). Knowledge sources and impacts on subsequent inventions: Do green technologies differ from non-green ones? Res. Policy, 49(2), 103901. Barbieri, N., Marzucchi, A., & Rizzo, U. (2023). Green technologies, interdependencies, and policy. JEEM, 118, 102791. Cassi, L. & Plunket, A. (2015). Research collaboration in co-inventor networks: combining closure, bridging and proximities. Reg. Stud, 49(6), 936–954. European Commission (2023). The european green deal industrial plan. Accessed: 2024-09-23. Graf, H. & Menter, M. (2022). Public research and the quality of inventions: the role and impact of entrepreneurial universities and regional network embeddedness. Small Bus. Econ., 58, 1187-1204. JRC (2022). Towards a green & digital future. Joint Research Centre (JRC), European Commission. Kalthaus, M. (2020). Knowledge recombination along the technology life cycle. JEE, 30(3), 643–704. Su, Z., Ahlstrom, D., Li, J., & Cheng, D. (2013). Knowledge creation capability, absorptive capacity, and product innovativeness. R&D Management, 43(5), 473–485. Wanzenböck, I., Rocchetta, S., Kim, K., & Kogler, D. F. (2024). The emergence of new regional technological specialisations: exploring the role of organisations and their technological network structure. Industry and Innovation, (pp. 1–27). Weitzman, M. L. (1998). Recombinant growth. QJE, 113(2), 331–360.

ABSTRACT. Abstract The global shift towards a low-carbon economy has intensified in recent years, driven by the urgency of addressing climate change and the international commitments set forth in the Paris Agreement. For China, the world’s largest emitter of greenhouse gases, transitioning away from fossil fuels is not only an environmental imperative but also a strategic economic challenge. This research focuses on the transition risks faced by entrepreneurs and businesses within the Greater Bay Area (GBA)—a key economic powerhouse in China—amid the country’s ambitious decarbonization goals. By employing a mixed-methods approach, this study aims to identify the primary risks posed by the shift from fossil fuels and propose macroeconomic policies to mitigate the adverse impacts on direct stakeholders, including firms and entrepreneurs in the GBA.

Introduction The Greater Bay Area, comprising major cities like Hong Kong, Shenzhen, and Guangzhou, represents a critical region in China's economic landscape, contributing over 11% of the national GDP. Its strategic importance in driving innovation and industrial growth positions it as a focal point in the national push toward sustainable development. However, the transition away from fossil fuels presents substantial risks, particularly for firms deeply embedded in carbon-intensive industries. These risks include stranded assets, disrupted supply chains, market volatility, and regulatory uncertainty.

The study explores these challenges within a global context, drawing comparisons with similar transitions in other regions, such as the European Union’s Green Deal, the United States’ Inflation Reduction Act, and renewable energy transitions in Southeast Asia. Lessons from these regions highlight the critical role of well-designed macroeconomic policies in cushioning the impact on businesses and ensuring a just transition.

Research Questions What are the primary transition risks for firms in the Greater Bay Area during China’s shift away from fossil fuels? How have other regions addressed similar challenges, and what lessons can be applied to the GBA? What macroeconomic policies can be implemented to mitigate these risks and support stakeholders during the transition?

Methodology This research employs a mixed-methods approach, combining qualitative and quantitative methods. Semi-structured interviews with stakeholders, including entrepreneurs, policymakers, and finance experts, provide insights into transition risks, complemented by case studies of firms in carbon-intensive industries. Quantitative analysis includes financial data evaluation, macroeconomic modeling to assess policy impacts, and surveys capturing stakeholder perceptions and readiness. Comparative analysis examines global examples, such as the EU’s Green Deal and the U.S. Inflation Reduction Act, for transferable lessons. Thematic analysis integrates qualitative findings with quantitative trends. This approach ensures a comprehensive understanding of risks and policy solutions tailored to the Greater Bay Area.

Findings Transition Risks in the Greater Bay Area Economic Disruption: Firms reliant on fossil fuels, such as manufacturing and logistics companies, face stranded assets and increased operational costs due to carbon taxes and stricter regulations. Policy Ambiguity: Uncertainty surrounding the timing and stringency of climate policies creates investment risks, deterring innovation and long-term planning. Supply Chain Vulnerability: Many firms in the GBA depend on fossil fuel-driven logistics and raw material supply chains. The transition could lead to cost escalation and operational delays. Labor Market Impacts: Workers in fossil fuel-dependent industries risk unemployment, while the demand for green skills remains unmet, exacerbating social inequalities.

Global Comparisons European Union: The EU’s Green Deal demonstrates the importance of a phased transition supported by subsidies for renewable energy and retraining programs for workers. However, the policy’s stringent timeline has posed challenges for industries with high carbon footprints.

United States: The Inflation Reduction Act highlights the efficacy of financial incentives, such as tax credits for clean energy projects, in mobilizing private investment. Nevertheless, it has faced criticism for insufficient support for displaced workers.

Southeast Asia: Nations like Indonesia and Vietnam offer lessons in balancing economic growth with decarbonization. Their gradual shift from coal to renewables underscores the significance of international funding and public-private partnerships.

Policy Recommendations for the Greater Bay Area Gradual Implementation of Climate Policies: A phased approach to decarbonization, with clear timelines and milestones, can reduce regulatory uncertainty and enable firms to adapt incrementally.

Financial Support Mechanisms: Establishing green financing schemes, such as low-interest loans and subsidies for renewable energy adoption, can alleviate financial pressures on businesses.

Reskilling and Workforce Development: Investing in education and training programs for workers can address labor market disruptions and prepare the workforce for emerging green industries.

Public-Private Partnerships: Encouraging collaboration between government bodies and private firms can foster innovation and scale up renewable energy projects. Regional Collaboration: Leveraging the GBA’s integration with Hong Kong and Macau can promote knowledge-sharing and resource optimization.

Implications and Conclusion

The findings underscore the critical need for well-coordinated policies to manage transition risks in the GBA. By learning from global best practices and tailoring them to the region’s unique economic and industrial context, policymakers can minimize the negative impacts on businesses while fostering a sustainable economic model. The study concludes that managing transition risks effectively will not only bolster China’s climate ambitions but also reinforce the GBA’s position as a global economic leader.

This research contributes to the broader discourse on sustainable development by offering a nuanced understanding of transition risks and mitigation strategies in one of the world’s most dynamic economic regions. It also serves as a template for other regions navigating similar challenges, emphasizing the importance of collaboration, innovation, and inclusivity in achieving a just transition.

ABSTRACT. Decarbonisation is one of the critical strategies for countries to embark on sustainable and inclusive development pathways. Green technologies are the main driver, broadly understood as those that reduce greenhouse gas emissions or remove (and store) these gases from the atmosphere. To gain a comprehensive understanding of knowledge generation in the green transition in Latin America, this paper examines the extent to which new specialisation in green technologies is more likely in Latin American countries with existing knowledge bases in green technologies, non-green technologies, or both. It also examines how international or local collaborative networks determine which countries produce new green knowledge. The paper builds on two main strands of literature. The first one connects the idea of technological space, whereby a country/region's technological specialization is more likely to develop new technologies in innovations related to its current knowledge base than unrelated innovations. The second one relies on recent studies on network collaborations on green and nongreen innovations (Capone et al, 2024). In this vein, this paper explores whether green technological specialization in Latin American countries is facilitated or constrained by the nature of the interrelationships between global and local actors that make up the collaboration network of technological knowledge. Our results suggest that green and non-green knowledge base positively affects Latin America's likelihood of specialising in green technologies. Our preliminary results would only seem to confirm that intra connections (within the country) in the invention network have a positive impact on the possibilities of green technological specialisation. We believe that part of this results are reflecting the heterogenities of the innovation processes within Latin America previously highlighted by Bianchi et al (2021). The next steps in the work will focus on the robustness of our results.

ABSTRACT. Summary: This paper investigates the impact of collaborating with foreign-educated researchers on Colombian non-mobile researchers’ agendas. Data from a centralized academic CV system, scholarship lists, and a publication repository are combined to assess whether collaborating with foreign-educated drives non-mobile researchers to broaden or narrow the scope of his or her research agenda. By applying a Propensity Score Matching strategy and an Event Study analysis, we find that non-mobile researchers collaborating with foreign-educated researchers become less diversified than their colleagues. The effect is amplified as non-mobile researchers collaborate with a greater number of foreign-educated peers.

Extended abstract:

Background and rationale

Pursuing training abroad has been an effective strategy for researchers looking to advance their careers (Netz et al., 2020). Extant evidence at the individual level has shown that those who move abroad for training (i.e., foreign-educated researchers) have, on average, higher productivity (Baruffaldi et al., 2020; Baruffaldi & Landoni, 2012; Petersen, 2018), visibility (Jonkers & Cruz-Castro, 2013), and social capital (Jonkers & Tijssen, 2008; Scellato et al., 2015) than those who never experience mobility (i.e., non-mobile researchers). This is largely due to the fact that while aborad foreign-educated researchers increase their human and social capital (Bozeman & Corley, 2004; Lee & Bozeman, 2005). Also, according to knowledge recombination theory, mobility allows researchers to access distant sources of knowledge, which is often more beneficial in sparking creativity than is local knowledge (Fleming, 2001; Franzoni et al., 2018).

While much of the focus of the literature has been on the impact of mobility at the individual level, less attention has been given to the extended effects of mobility on non-mobile. Drawing on the argument that knowledge and social capital from foreign-educated researchers can be absorbed by other researchers through formal collaboration, or through being located at the same department, scholars have provided evidence that foreign-educated researchers positively influence nonmobile researchers’ productivity and networks (Borjas & Doran, 2015a; Franzoni et al., 2018; Fry & Furman, 2023; Fry, 2023; Müller et al., 2023; Yang et al., 2022). However, among the questions that remain unaddressed, we ask the following: since foreign-educated researchers access different sources of knowledge, does collaboration with foreign-educated researchers help non-mobile researchers expand their research agendas?

Methodology

In this paper, we assemble a unique dataset on Colombian researchers by combining three data sources: the national academic CV repository, Ph.D. scholarship lists, and OpenAlex, using observations from 1990 to 2021. We reconstruct researchers’ educational profiles and work trajectories, identifying foreign-educated and non-mobile researchers. To assess the impact of collaborating with a foreign-educated researcher on a non-mobile scientist’s research agenda, we consider two diversification indicators: (i) we calculate the share of papers published after having started the collaboration in which 75% of their (OpenAlex defined) topics are not present in their previous publications; (ii) we calculate a “pivot index” to capture the directional change of research. For the latter indicator, we modify the pivot index measure proposed by Hill et al. (forthcoming) by using topics assigned to publications in OpenAlex instead of referenced journals from a researcher’s publications. The methodological approach combines a matching and difference-in-differences estimator, relying on nearest-neighbor matching to find comparable treatment and control groups. We then rely on Callaway & Sant’Anna’s (2021) difference-in-differences estimator to perform an Event Study and obtain the Average Treatment Effect on the Treated (ATT). Results

We find that collaboration with foreign-educated researchers reduces the diversification of non-mobile researchers with respect to the other Colombian researchers. The share of publications of non-mobile researchers who co-author with foreign-educated researchers having 75% new topics is 18 percentage points lower than that of the control group. This indicates that non-mobile researchers collaborating with foreign-educated researchers are more inclined to publish on topics they are already familiar with. Regarding the pivot index, we observe an 11-percentage point decrease in the research direction trajectory of the non-mobile researchers. Overall, these results indicate that, for non-mobile researchers, specializing in fewer topics is the result of collaborating with foreign-educated researchers. Furthermore, by conducting an Event Study, we observe a consistent decrease in the share of publications with new topics and the direction of research trajectory (measured by the pivot index) over the years as non-mobile researchers increase their stock of foreign-trained coauthors.

Significance

We contribute to the literature on international scientific mobility in two ways. First, we analyze the extended effects of international mobility. Along with a few scholars (Fry & Furman, 2023; Fry, 2023; Müller et al., 2023), we depart from the mainstream literature, which emphasizes the impact of mobility on those who experience it, to demonstrate that an international mobility experience can also influence individuals who have never moved abroad. Second, we advance the literature by focusing on the extended effects on non-mobile researchers’ agendas. Previous studies have primarily focused on the effects of productivity and social capital. However, the impact of collaborating with foreign-educated researchers on the topics of non-mobile research remains barely explored. In our case, we show that by collaborating with foreign-educated, non-mobile researchers are stimulated to focus on topics they are familiar with. This seems likely to suggest that foreign-educated researchers play a role in helping locals to specialize, leveraging the absorptive capacity matured in the topics they know to refine their skills and include new methods or applications in their research.

ABSTRACT. Over the past decades, international research collaboration has grown significantly, reflecting its essential role in advancing knowledge and addressing complex global challenges. This trend is especially prominent in the life sciences and medicine, where collaborative efforts are critical to tackling issues such as global health crises, disease prevention, and the improvement of basic human well-being. Researchers working across national borders often combine complementary expertise, resources, and perspectives, creating opportunities to address multifaceted problems that transcend geographic and disciplinary boundaries (Wagner et al., 2015). Despite the recognized importance of these collaborations, the micro-mechanisms shaping them remain underexplored, particularly in contexts marked by differences in available resources and scientific capacity among collaborators.

This study aims to investigate the factors influencing research collaboration between researchers from diverse economic contexts, focusing on differences in motivations, perceptions of research work, and the challenges faced during collaborations. To provide a nuanced understanding of these dynamics, the study also accounts for prior collaboration experience and researcher mobility. By integrating data from large-scale surveys and analyzing the specific experiences of researchers involved in life sciences, the study contributes to an evidence-based understanding of international research partnerships, especially in addressing key gaps in existing literature.

International collaboration has been frequently promoted through policy initiatives and is seen as a mean to enhance scientific capacity and foster global scientific exchange (Bradley, 2008). These collaborations are particularly vital in the life sciences, where advances often depend on access to diverse expertise, cutting-edge technologies, and data-sharing frameworks (Carvalho et al., 2023; Matenga et al., 2019). However, the global distribution of scientific capital and resources remains uneven, with countries in the Global South often facing resource constraints that can shape the dynamics of their participation in international collaborations (Fonseca et al., 2018; Mbaye et al., 2019).

While disparities in resources and capacities have been a key focus in studies of international research collaboration, less attention has been paid to the motivations and practical experiences of researchers themselves. Prior studies, such as those by González-Alcaide et al. (2017) and Gonzalez-Brambila et al. (2013), have primarily employed large-scale quantitative approaches to analyze collaboration networks and bibliometric indicators on the macro-level, while others have relied on case studies and qualitative interviews to explore specific contexts and partnerships (Dusdal & Powell, 2021; Kontinen & Nguyahambi, 2020). These approaches have highlighted a broad spectrum of collaboration experiences, ranging from highly successful and mutually beneficial partnerships to relationships perceived as exploitative or asymmetrical (Gunasekara, 2020; Munung et al., 2017).

Our study bridges this gap by employing a survey-based approach to examine the motivations, barriers, and challenges in collaborations between researchers in different economic contexts. Specifically, it seeks to understand how researchers perceive and navigate the complexities of international collaboration, with a focus on life sciences and related fields. Previous survey studies on international collaboration, such as those by Matthews et al. (2020) and Muriithi et al. (2018), have identified key factors influencing collaboration decisions, including resource dependence, the availability of funding, and institutional policies. Building on these findings, our study examines a broader range of factors, including intrinsic motivations, access to expertise, and the impacts of collaboration on researchers’ careers.

The study is based on an online survey distributed to 6,646 researchers in the life sciences, resulting in 1,540 responses (a 23% response rate). The participants were selected from authors of articles published in prominent life science journals between 2018 and 2022, focusing on papers with no more than five authors, as identified through Scopus. In the first round, corresponding authors of selected publications in these journals were invited to participate. In the second round, co-authors of responding corresponding authors were invited to participate, allowing for triangulation and comparison of perceptions among collaborators working on the same research projects.

The personalized survey addressed various aspects of researchers’ experiences with collaboration, including project characteristics (e.g., funding sources, project duration, and collaboration initiation), demographic information of respondents and their co-authors (e.g., affiliations, gender, academic position, and mobility history), and perceived contributions to the project. Participants were also asked about their motivations for collaboration, challenges encountered, and the outcomes of these partnerships.

Preliminary descriptive findings indicate that while disparities in scientific resources may shape certain aspects of collaboration, many motivations and challenges are shared across contexts. Intrinsic motivations, such as personal interest and intellectual curiosity, consistently emerged as the most significant factors shaping research questions. Among respondents, personal interest ranked higher than considerations such as collaborators’ preferences, funding requirements, data availability, or publication opportunities. This finding highlights the central role of individual agency and intellectual engagement in driving research efforts, even within the context of international collaborations.

When selecting collaborators, participants across economic contexts consistently ranked access to complementary knowledge and skills as the most critical motivator. This factor was perceived as more important than other practical considerations, such as access to funding, infrastructure, data, or increased visibility of research outcomes. This underscores the importance of aligning expertise and capabilities within research teams, which can enhance the quality and impact of collaborative projects.

Despite these shared motivations, some challenges remain context-specific. Sourcing funding emerged as the most significant obstacle, particularly for researchers working with collaborators in resource-constrained settings. This finding aligns with prior research highlighting the uneven distribution of research funding and its implications for international collaboration (Sridhar, 2012).

These early findings contribute to a more nuanced understanding of the factors influencing international research collaboration. They highlight both the shared motivations that drive researchers to engage in collaborative work and the specific challenges that can arise in resource-constrained contexts. By focusing on the lived experiences of researchers, this study provides valuable insights into the dynamics of collaboration and the mechanisms that support or hinder productive partnerships.

Further analyses of the data are under way with the objective of deepening understanding of these patterns and explore their implications for policy and practice. These insights can inform strategies to foster global scientific exchange and address critical challenges in the life sciences, ultimately enhancing the collective ability to tackle pressing global issues.

ABSTRACT. This paper’s title paraphrases Robert Axelrod’s best seller “The Evolution of Cooperation”. Instead of focusing on how international cooperation emerges and develops, we aim at understanding why cooperation does not evolve. Departing from International Political Economy perspectives, we question to what extent the global economic position of countries in certain sectors explains the absence or insufficient regulatory measures adopted by states through the design and implementation of international agreements. We focus on the case of bioprospecting in Antarctica, a practice that has emerged as a new form of exploitation of the continent's resources between the late 1980s and early 1990s. Though Antarctica has historically been a source of economic resources and disputes, the Antarctic Treaty, signed in 1959, has not accounted for economic issues. Even if, since then, the trajectory of the Antarctic regime has been characterized by the evolution of cooperation in some economic-related areas, bioprospecting has not been part of Antarctic binding instruments. While the vast majority of the Antarctic Treaty's consultative members, at different degrees, have sustained a cooperative position by offering information and feeding debates about bioprospecting practices during Consultative Meetings, the fact that substantial decisions demand consensus could mean that a minority possesses a veto power and exercises it by remaining silent. In order to understand to what extent economic interests have influenced cooperative and non-cooperative positions adopted by consultative members of the Antarctic Treaty we have departed from two hypotheses: that governmental decisions which contribute to keeping the topic of bioprospecting out of the agenda of Antarctic regulations are influenced by the economic preferences of countries that are well positioned as global technological and commercial players; that cooperation can be promoted by players that aim at improving their global relative economic position. We adopt the geographic distribution of Antarctic simple patent families filed until 2023 as a proxy for global competitiveness. The number of bioprospecting-related working, information, and background papers (which we call 'position papers') submitted by governments to consultative meetings until 2023 is adopted as a proxy for cooperativeness. We recognize that the best indicator of economic competitiveness is trade. However, as there is no comprehensive database on Antarctic biotechnology-related traded products we have relied on a Boolean research in Lens.org that combined the list of OECD's biotechnology-related patent codes with Antarctic-related terms. Through a bidimensional analysis the mean of both variables was calculated in order to distribute the analyzed countries in four quadrants, considering the number of positioning papers and of patents above and below the mean. As the US has been identified as an outlier (with 1240 patents, followed by Japan, with 133 patents) it was removed from the dataset for mean calculation. Our hypothesis on the relationship between economic competitiveness and non cooperation seems to hold mainly in the case of Antarctic patent leaders that did not submit any position papers - cases of the US, Japan, China and South Korea. As hosts of powerful Antarctic innovation actors, those four silent countries might hinder movements that can lead to the regulation of their economic activities (as they prosper in laissez-faire) and to sharing strategic information that can enhance the know-how of potential competitors. We also found that other Antarctic patent leaders, such as the UK, Germany, Norway and Australia, have adopted lower cooperative positions. Higher cooperative positions were found in the cases of the Netherlands (15 position papers) and Belgium (11 position papers). Though both do not host many inventors of Antarctic biotechnology related patents, they are well positioned in the biotechnology-based economy in general and thus could benefit from cooperation initiatives that promote transparency and access to information related to Antarctic activities developed by current leaders. We suggest caution in treating both cases as similar to other countries that also appeared in the same quadrant, such as Argentina and Brazil, which, not being reference countries in biotechnology, may not have enough absorptive capacities to benefit from more access to information that can be promoted by cooperative arrangements. The case of France, the only country that appears in one of the quadrants, is also worth mentioning, as it can be considered an Antarctic patent leader, but, having issued six position papers, also seems to hold a more cooperative position towards bioprospection. As a country that maintains territorial claims in Antarctica - also the case of Argentina, Australia, Chile, New Zealand, Norway, and the UK - France can feed debates on Antarctic bioprospecting in an attempt to induce reciprocity and promote transparency from countries that have been bioprospecting in the Antarctic territory it claims. We conclude by stressing that both cooperation and non cooperation, as political economy phenomena, have distributional consequences. The non-regulation of bioprospecting activities in Antarctica benefits current leaders, as its regulation can benefit potential leaders, but not countries that do not have the indigenous networks needed to become leaders. For future research we recommend that the role of non-consultative members, such as Canada and Denmark, is accounted for, since they seem to be important Antarctic innovators. We recognize that decision-making in international negotiations is a complex process that involves not only economic interests, but also their interaction with other interest groups and institutions, including social movements, scientific organizations and the military. Therefore the next stage of our research is focused on case studies. Qualitative analysis will be crucial to allow us to hypothesize on the interaction between economic interests and other variables that can affect cooperation.

ABSTRACT. Scientific research funding plays a crucial role in fostering talent and advancing innovation across various disciplines. By providing financial support to researchers, these funds aim to promote academic productivity, career development, and collaborative networks. However, the general mechanisms and impacts of such funding on talent development require further exploration. This study examines the broad influence of scientific research funding on the growth and career trajectories of researchers, focusing on its effects on research output, professional advancement, and international collaboration. Research Questions: This study addresses the following questions: (1) How does scientific research funding influence researchers' academic productivity and career progression? (2) What are the short-term and long-term effects of funding across different disciplines and career stages? (3) How do funded researchers compare with non-funded peers regarding professional growth and academic influence? Research Methods: A quasi-experimental design is adopted, utilizing difference-in-differences (DiD) and propensity score matching (PSM) techniques. The analysis draws on data related to research publications, citation metrics, career promotions, and collaboration networks. Funded researchers are compared to a matched control group of non-funded researchers to isolate the causal effects of funding on their development. Research Design: This study uses recipients of the National Natural Science Foundation of China (NSFC) Excellent Young Scientists Fund (EYSF) between 2013 and 2018 as the primary sample. The primary growth metric is whether recipients subsequently obtain the Distinguished Young Scientists Fund (DYSF), and the time taken to transition from EYSF to DYSF is analyzed as an indicator of career advancement. Key academic outcomes are measured using metrics such as the total number of publications, citations, and H-index, providing a quantitative assessment of academic productivity and impact. Furthermore, a co-authorship network is constructed based on the publications of the sampled researchers to evaluate their collaborative relationships and the development of academic networks over time. Propensity score matching (PSM) is applied to select a control group of non-funded researchers with similar academic and professional backgrounds. A difference-in-differences (DiD) approach is used to compare the trajectories of funded and non-funded researchers, with multivariate regression models controlling for discipline, institutional type, and demographic factors. This design aims to provide a comprehensive understanding of the impact of scientific research funding on talent development, focusing on both individual growth and collaborative innovation. Expected Outcomes: The study is expected to reveal that scientific research funding significantly boosts academic productivity, enhances professional advancement, and fosters international collaboration. These effects may vary by discipline and career stage, highlighting the need for tailored funding strategies. The findings will offer evidence-based recommendations to optimize funding mechanisms for supporting talent cultivation and achieving broader research and innovation goals.

ABSTRACT. Environmental regulation experiences one of the most critical challenges in methods of regulation in the contemporary world due to the inability of conventional non-dynamic regulatory approaches to address dynamic ecology factors, constantly transforming industries, and fluctuating resource consumption trends. The present study focuses on embedded AI systems in environmental policy, applying a new approach to developing intelligent and flexible environment policies for the effective management of resources. To this end, our study contributes to the existing literature on environmental policy and practice by identifying the following important lacunas in the existing rature: Examining how AI-regulatory tools can be implemented to foster accountability, transparency, and ecological sustainability still remains a major gap in the present literature.

The research utilises a novel methodological approach with machine learning, environmental science, and policy to introduce the concept of Smart Environmental Regulation (SER). It is also important to note that our approach aggregates real-time data from EMSs, satellite imagery, and IoT sensor networks alongside compliance records of three geographically diverse locations to generate a coherent dataset for AI model training and validation. The research established that it would be possible to increase the effectiveness of regulations and decrease compliance expenses and environmental impacts using sophisticated algorithms for machine learning.

Our methodology encompasses four interconnected components: The four objectives mentioned in the paper includes: (1) proposing the development of an adaptive AI framework that engulfs real-time environmental data to disseminate necessary information to the regulators; (2) proposing the designing of a dynamic policy optimization system that would sequentially adjust regulatory parameters as per environmental conditions and the compliance history; (3) proposing the development of a transparent decision support system for the policymakers; and (4) proposing that which explains how accountability

Key findings demonstrate significant improvements in both regulatory effectiveness and resource management efficiency:

1. The created AI-driven regulatory system increased the chances of early identification of the non-compliance issues by the environmental regulators by 42% than the conventional monitoring techniques. 2. In this case, the use of adaptive resource management procedures enabled the firm to cut its response time to environmental incidences by 31% while enhancing resource utilisation by 28%. 3. The adaptive regulatory framework was also shown for a 35% improvement in enforcement actions with a 44% improvement in false positives. 4. A recent CK poll revealed that the public endorsed the application of artificial intelligence to regulate industry by 76 percent, based on the enhanced measurability of enforcement activities and increased hearing transparency.

Our research responds to several important questions in environmental regulation, including quality of data for environmental regulation, interpretability of the models, and environmental justice. We also offer new ideas embracing an empirically confirmed framework for environmental data accuracy along with an original approach to algorithmic fairness in terms of regulation enforcement. The study also introduces a new metric for evaluating the effectiveness of AI-driven environmental regulations: To supplement the DRI, I propose a second measure of regulatory ecosystems, the Dynamic Environmental Response Index (DERI), which measures the ways in which regulatory systems can respond to changes in environmental conditions.

Another substantial benefit of this work can be identified in the elaboration of the conceptual basis for an integrated policy approach to the use of artificial intelligence in the sphere of environmental regulation. This framework responds to the main potential issues of automation of regulatory activities but retains human intervention as necessary. This work shows how AI can support instead of replace human decisions concerning environmental policies, offering tools to make people more effective, consistent, and rational in decision-making.

The study reveals several transformative applications of AI in environmental regulation:

1. Predictive Compliance Monitoring: That is why AI systems learn about patterns in environmental data, containing the data that can help prevent compliance problems before they occur, which will allow exercising regulatory functions preventively. 2. Dynamic Resource Allocation: ML solutions allow for the more efficient allocation of regulatory assets with respect to the real-time evaluation of the risks or environmental effects. 3. Adaptive Policy Implementation: Since AI-based systems involve learning from the environment, they adapt the regulatory parameters in relation to policy objectives while conforming to environmental changes. 4. Environmental Justice Integration: The structure encompasses social and demographic information so that no one community gets preferential treatment in regards to environmental ordinances.

Our research shows that the potential for the increased application of AI in environmental regulation is coherent and suggests that it may lead to making resource regulation more effective and efficient. The study establishes that resource conservation regulatory outcomes are 37% higher when AI techniques are used as opposed to conventional approaches. Also, the study demonstrates how machine learning can refine relationships within the environment that might be difficult to capture with other forms of regulation.

The study also brings into discussion the general issues about environmental policymaking and how AI could close the perception gap. We find out that policy-oriented regulatory systems facilitated by AI are more likely to make quicker adjustments to changing environments and scientific discoveries than lose track of policy goals.

Future research directions identified in this study include the need for: - Improved implementation of the corresponding regulatory standards with the use of citizen science data by AI - Promotion of the norms for the use of AI to evaluate environmental effects - Research on best practices in the cooperation of environmental regulation between states - New techniques for employing Traditional Ecological Knowledge in Artificial Intelligence systems

This research adds to the growing body of literature on smart environmental regulation while offering a roadmap for policymakers and regulatory institutions. The results show that AI could improve the delivery of environmental policy as well as mitigate associated structural issues in resource management.

ABSTRACT. Green hydrogen is a promising pathway to reconcile economic growth and environmental sustainability. On the latter, green hydrogen offers significant potential to decarbonise hard-to-abate sectors such as steel and chemical industries, contributing to broaden the energy transition. On the economic side, the academic literature states that green hydrogen development enables the creation of new industries, jobs, and technological learning, among others. Therefore, green hydrogen is an interesting development alternative for low- and middle-income countries like Kenya, Uruguay, Namibia, India, Brazil. Chile, China, and South Africa, among others, who are developing hydrogen strategies and projects. However, as taught by the experience with other renewables, grasping economic benefits depends on a complex and dynamic interplay of domestic capabilities, resource endowments, industrial policy implementation, and political economy factors. Currently, there is little evidence and knowledge of the strategies followed by countries entering the nascent green hydrogen sector and we do not know the elements that will be conducive to better outputs and challenges faced by countries both because of the early development stage of this market sector but also because we lack systematic explorations of the strategies and experiences of early movers in green hydrogen in the Global South. Thus, countries in the Global South that are already moving into green hydrogen offer an invaluable learning opportunity for other low- and middle-income countries (LMIC) that are entering this market as they it offers insights on the approaches and policy options. We compile a unique dataset by combining secondary data from interviews with key stakeholders, official governmental documents, and academic and grey literature. We use it to do a use comparative case study and examine the green hydrogen approaches of Brazil, Chile, China, and South Africa. We show that their hydrogen strategies diverge considerably due to their distinct resource endowments, energy infrastructures, and market structures. Considering the public policies and production projects taking place, again the approaches vary in terms of actors engaged and industrial policies in place, but responses not always reflect the strategies devised initially. This highlights the of comprehensive cross-country exploitations of the topic as opposed to case studies of single countries as prevalent in the literature so far.

ABSTRACT. Abstract

The global energy landscape is undergoing a profound transformation as the world seeks to address the dual imperatives of achieving universal energy access and mitigating climate change. As nations and corporations set increasingly ambitious targets to reduce greenhouse gas emissions, the concept of net-zero frameworks has emerged as a central element of energy policy. Simultaneously, the Fourth Industrial Revolution (4IR) characterized by the integration of advanced technologies such as artificial intelligence (AI), the Internet of Things (IoT), blockchain, and big data has unlocked new possibilities for improving energy efficiency, reliability, and accessibility. While the promises of these innovations are immense, their potential to enhance energy access, particularly in underserved regions, remains an area of active inquiry and debate. This research seeks to explore how 4IR technologies can drive energy access enhancements while aligning with global net-zero frameworks, ultimately contributing to sustainable development and climate resilience.

The existing literature provides valuable insights into the role of 4IR technologies in transforming energy systems and the importance of net-zero frameworks in addressing climate change. However, there is a notable gap in research that directly examines the intersection of these two areas in the context of energy access. Most studies tend to focus on either technological innovations or net-zero policies in isolation, without considering how they can be integrated to address energy poverty. While some research has explored the potential of 4IR technologies in improving energy access in remote areas, few studies have systematically analyzed how these innovations can be aligned with net-zero frameworks to provide sustainable and inclusive energy solutions. Furthermore, existing studies often fail to incorporate a socio-technical systems approach, which is essential for understanding the complex interplay between technology, society, and policy in shaping energy outcomes. This research intends to fill this gap by examining how 4IR innovations, when integrated with net-zero frameworks, can enhance energy access. It also seeks to explore the potential synergies and tensions between the rapid adoption of emerging technologies and the long-term goals of net-zero emissions, particularly in the context of equity and inclusivity. By providing a comprehensive analysis of these dynamics, the research aims to contribute to the ongoing discourse on sustainable energy transitions and inform policy decisions that support both climate action and social equity.

To achieve these objectives, the research builds on the socio-technical systems theory. This examines how technological innovations interact with social, economic, and institutional structures to create systemic change. In the context of this research, the adoption of 4IR technologies in energy systems can be analyzed as part of a broader transition in socio-technical regimes, at the same time exploring the interdependencies between technological innovations (e.g., AI, IoT), societal needs (e.g., energy access), and institutional frameworks (e.g., net-zero policies).

This research employs a combination of secondary data analysis, case study evaluation, and qualitative content analysis to derive insights into the theoretical and practical aspects of energy access, technological innovation, and sustainability transitions. These case studies examine a variety of contexts, from large-scale renewable energy projects in developed economies to decentralized, off-grid solutions in rural areas of the Global South, with particular attention to the challenges posed by regulatory frameworks, financing mechanisms, and social acceptance. Qualitative content analysis of relevant literature and policy documents is used to extract key themes and insights related to the integration of 4IR innovations with net-zero frameworks. This analysis helps identify recurring patterns and challenges across different regions, as well as gaps in current knowledge and policy implementation. The data from these sources is analyzed through thematic and comparative analysis, allowing the research to draw broader conclusions about the effectiveness of different technological and policy approaches.

This research contributes to the growing body of literature on the role of technology in energy transitions, with a specific focus on how emerging innovations can drive social and environmental change. By exploring the intersection of 4IR technologies and net-zero frameworks, the study provides a deeper understanding of how these two forces are harnessed to address the pressing challenge of energy access. One of the key contributions of this research is its examination of the synergies and tensions between technological innovation and policy frameworks. While 4IR innovations hold great promise for enhancing energy access, their widespread deployment must be aligned with the goals of net-zero emissions and sustainable development. This alignment requires the development of innovative policy instruments that can foster collaboration between governments, private sector actors, and civil society, ensuring that energy transitions are both technologically feasible and socially inclusive.

In addition to its contributions to academic literature, this research has significant policy implications. For example, while net-zero frameworks are often focused on large-scale transitions in the electricity grid, 4IR innovations can offer pathways to decentralized, localized energy solutions that are better suited to the needs of rural communities and off-grid populations. Policymakers will need to consider how to integrate these innovations into existing infrastructure and regulatory systems, while also addressing the social and economic factors that influence energy access.

In conclusion, this research provides valuable insights into how 4IR innovations and net-zero frameworks can work together to enhance energy access, particularly in developing economies. By identifying the key drivers and challenges of this intersection, the study offers concrete policy recommendations that can support inclusive and sustainable energy transitions worldwide. Ultimately, this research aims to contribute to the broader goal of achieving universal access to modern energy services while mitigating the impacts of climate change in line with the global commitment to a net-zero future.

ABSTRACT. The growing ubiquity and sophistication of algorithms have transformed industries and economies worldwide, particularly within the digital economy. Algorithms drive innovation and operational efficiency, enabling the emergence of platform economies that bridge suppliers and consumers through data-driven matchmaking and management systems. Ride-hailing platforms, in particular, rely on algorithms to optimize operations, from dispatching drivers to evaluating worker performance. While these algorithmic systems enhance productivity and reduce transaction costs, they also introduce significant challenges. The opacity and complexity of algorithmic decisions often undermine worker autonomy and exacerbate power asymmetries between platforms and workers. This imbalance raises critical questions about fairness, accountability, and transparency in algorithmic management. Furthermore, traditional regulatory frameworks struggle to adapt to the unique challenges posed by algorithm-driven economies, necessitating innovative approaches to algorithmic governance. Despite a growing body of research on algorithmic management, little is known about how to design regulatory systems that balance efficiency with fairness while addressing the societal impacts of algorithmic decisions. This study employs a simulation-based approach to analyze the implications of algorithmic strategies in ride-hailing platforms, providing actionable insights into algorithm regulation and platform governance.

The research utilizes an agent-based simulation model, constructed on a Markov Decision Process (MDP) framework, to explore the dynamics of algorithmic decision-making within a two-sided ride-hailing market. The model examines how algorithms balance efficiency and fairness in the context of platform operations, focusing on key metrics such as driver income, income distribution fairness, passenger waiting times, and platform profitability. The MDP framework defines state variables, including the number of active drivers and passengers as well as average driver working hours, and models platform actions as trade-offs between prioritizing proximity in dispatching and promoting income fairness among drivers. The reward function evaluates outcomes based on system performance, enabling the simulation to assess the long-term impacts of different algorithmic priorities. Two optimization algorithms, Proximal Policy Optimization (PPO) and Deep Q-Networks (DQN), are employed to simulate the evolution of platform strategies under varying objectives, such as maximizing profitability or ensuring fairness. The simulation iteratively updates the market state based on interactions between drivers and passengers, offering a detailed analysis of the trade-offs involved in algorithmic management. By systematically varying simulation parameters, the study evaluates the robustness of its findings and explores the potential impacts of platform decisions on various stakeholders.

The findings reveal the intricate trade-offs inherent in algorithmic management. Efficiency-focused strategies, which prioritize passenger wait times and platform profitability, often lead to disproportionately low wages and extended working hours for drivers. Conversely, fairness-focused strategies improve income equity among drivers but result in longer passenger waiting times, potentially undermining user satisfaction and platform growth. These results highlight the challenges of simultaneously achieving high driver income, low passenger wait times, and maximum platform profitability. Furthermore, the study underscores the critical role of transparency in algorithmic governance. The opacity of algorithmic decision-making processes not only exacerbates power imbalances between platforms and workers but also erodes trust and fairness in the marketplace. Enhancing algorithm transparency and accountability is thus essential to fostering equitable outcomes and restoring stakeholder trust in platform economies.

The implications of these findings extend to policymakers and platform operators, underscoring the need for adaptive governance frameworks that balance efficiency and fairness. Policymakers should establish regulatory standards that promote algorithmic transparency and fairness while encouraging platforms to adopt balanced strategies that consider broader societal impacts. Collaborative governance, involving regulators, platforms, and stakeholders, is crucial to developing innovative solutions that address the inherent trade-offs of algorithmic management. Additionally, platforms must prioritize worker empowerment, providing tools and resources that allow workers to better understand and influence the algorithms that shape their livelihoods. These measures can help ensure that algorithmic systems serve the interests of all stakeholders, contributing to a more equitable and sustainable digital economy.

This study provides a critical contribution to the ongoing discourse on algorithmic governance, highlighting the need for innovative regulatory approaches that align technological advancements with social welfare. By integrating simulation-based analysis with practical policy recommendations, the research lays the groundwork for a balanced approach to algorithmic management in platform economies, offering valuable insights for future research, policymaking, and platform governance. The findings emphasize the importance of transparency, equity, and accountability in algorithm design and implementation, advancing the broader goal of achieving sustainable and equitable outcomes in the digital economy.