ABSTRACT. AI-based Decision Support Systems (AI-DSSs) can support complex decision-making in supply chain and manufacturing, but their practical value depends on whether users can appropriately accept, question, or override AI advice. This doctoral research proposal represents a research agenda on examining advice acceptance in Human-in-the–Loop AI-DSS under the Judge–Advisor System paradigm, where AI provides recommendations and humans retain final authority. It argues that existing research often studies mechanisms such as explainability and interactive control in isolation, offering limited guidance on how they should be configured for different decision tasks. The proposed research addresses this gap by developing a task-sensitive framework that links supply chain and manufacturing task characteristics with suitable AI-DSS mechanisms and interactive features. The proposal outlines a systematic review and a planned evaluation with business experts as steps toward more dependable organizational AI-DSS.

ABSTRACT. AI-supported advisory systems are moving into customer-facing financial services. In insurance, their acceptance is especially difficult to explain because advisory decisions concern complex products, future risks, and situations in which customers cannot fully evaluate recommendation quality on their own. This doctoral project examines the acceptance of AI-supported insurance dvice in the German insurance market. Building on the Technology Acceptance Model, the dissertation develops a focused model in which trust mediates the relationship between perceived usefulness, perceived ease of use, and intention to use AI-supported insurance advice. The planned empirical design is a scenario-based online survey with end customers, followed by structural equation modeling. At the current stage, the model and research design are deliberately kept open to further refinement. Feedback is particularly sought on the positioning of the trust construct, the adequacy of a scenario-based design, and the balance between a parsimonious core model and context-specific extensions.

ABSTRACT. Fairness and transparency are increasingly important requirements in collective decision-making processes involving limited resources and conflicting stakeholder interests. Real-world allocation problems such as workplace scheduling, hot-desking, meeting room assignment, or shift planning are often computationally complex and require specialized combinatorial optimization techniques. While recent approaches like SocialCOP enable the integration of formal fairness criteria into optimization models, the resulting decisions frequently remain difficult to understand and accept for non-technical stakeholders. This research proposes a participatory and explainable framework for fair combinatorial optimization that combines constraint programming, explainable AI techniques, and large language models (LLMs). The goal is to develop adaptive optimization systems that allow stakeholders to actively explore, modify, and negotiate fairness constraints and trade-offs within complex allocation problems. The project extends existing SocialCOP approaches with explainability mechanisms such as minimal unsatisfiable subsets, minimal correction sets, and counterfactual explanations to make optimization decisions transparent and interactive. Furthermore, the research investigates how natural-language interfaces and LLM-supported explanations can facilitate accessibility for users without technical expertise while preserving the correctness guarantees of formal optimization systems. Through participatory workshops, user studies, and field experiments, the project evaluates both measurable fairness metrics and subjective perceptions of fairness, trust, and acceptance. The expected outcome is a human-centered AI framework that enables socially acceptable, explainable, and adaptable decision support for complex combinatorial optimization problems.

ABSTRACT. In this talk, we will consider the bits and bytes of writing a research paper. We will look at the design, outline, and different aspects of writing a paper, as well as related aspects such as reviews, ethics, collaboration, embedding in a larger context, publication strategy, etc. During and after the talk, there will be room for questions and comments.

ABSTRACT. Large Language Model (LLM)-based multi-agent systems (MAS) are increasingly used for distributed problem-solving tasks. However, these systems exhibit persistent failure modes, including coordination breakdowns, inconsistency across interactions, and vulnerability to adversarial manipulation. This work argues that these failures are strongly linked to epistemic limitations, specifically the lack of explicit mechanisms to represent and reason about agents’ knowledge and beliefs. To address this, this research proposes a neuro-symbolic framework that introduces an explicit epistemic layer for multi-agent coordination. The framework models agents’ beliefs, knowledge dependencies, and uncertainty as a structured epistemic state, enabling reasoning about knowledge alignment and inconsistency. In addition, epistemic reasoning is positioned as a form of structured context engineering for long-horizon interactions. The approach is evaluated through targeted experiments focusing on coordination stability, epistemic reasoning, and adversarial robustness, aiming to demonstrate improved consistency and robustness in distributed MAS.

ABSTRACT. The integration of Large Language Models (LLMs) into educational environments enables new forms of automated content generation, including microlearning units derived from course materials, textbooks, and lecture notes. However, generating pedagogically sound and source-faithful learning content often requires extensive prompt engineering, and this challenge becomes even more relevant when privacy constraints prevent the use of large closed-source models. This dissertation addresses this problem by developing a reference-free Automated Prompt Optimization (APO) pipeline for improving LLM-generated microlearning units. The proposed system combines Retrieval-Augmented Generation (RAG), human-aligned LLM-as-a-Judges (LaaJs), and reflective prompt optimization. The creation system generates self-contained H5P microlearning units, while the evaluation system provides criterion-based textual and scalar feedback. This feedback is then used to optimize the prompts of the creation system without requiring ideal pre-authored reference units. The pipeline is model-agnostic and can be used with closed-source models of choice as well as locally deployed models. The empirical focus, however, lies on resource-efficient current open models such as Gemma 4 26B A4B, while larger models remain optional for environments where privacy, cost, and infrastructure constraints allow their use.

ABSTRACT. Rule-based systems and machine learning (ML) are two different approaches to model and solve a variety of problems. Rule-based systems are reliable and interpretable, but often lack flexibility and scalability. ML models are flexible and scalable, but can be unreliable and difficult to interpret. The hybridization of rule-based systems and ML is a promising approach to profit from the advantages of both methods. The aim of this work is to examine model-agnostic possibilities to automatically extract interpretable rules, especially constraints, out of machine learned models to handle situations where no knowledge-base is given and constraints cannot be implemented right away. The study specifically examines data deduplication as an application domain, where transportation data often contain duplicate records due to multiple entries with slight variations by different contractors. The research questions focus on the techniques for rule extraction, formalization of rules into constraints, feature extraction, and the effectiveness of the extracted rules compared to traditional approaches.

ABSTRACT. Multimodal learning often suffers from modality imbalance, where some modalities dominate optimization while others remain under-optimized. This proposal studies this problem from the perspective of local preference inconsistency. We hypothesize that multimodal conflicts arise when different predictions within the same model agree on the target but disagree on which competing outputs are most plausible around it. We investigate this hypothesis in two settings. First, in imbalanced multimodal classification, we study inconsistency between unimodal branches and the fusion prediction, and explore how such inconsistency can guide modality-specific supervision. Second, in multimodal continual instruction tuning under MoE-LoRA, we study whether expert-local token preferences diverge from the model-global prediction and contribute to forgetting across sequential tasks. The goal is to develop sample-level mechanisms that preserve useful specialization while reducing destructive cross-component interference.

ABSTRACT. The circular economy is one main pillar for building a sustainable future. For that products need to be designed based on eco-design principles. Life Cycle Assessment (LCA) is needed to support eco-design by evaluating environmental impacts. However, data scarcity is a major challenge in LCA studies. To address this issue, data imputation and matrix completion techniques can be used to fill in missing data and improve analysis accuracy. However, a practicioner has the choice between many different algorithms which might or might not fit to the properties of the data at hand. Thus, a recommender system is proposed that provides recommendations which matrix completion algorithms to use depending on the use case. Furthermore, by studying the energy related impact of each algorithm, the tool also gives recommendations on the expected trade-off between performance and energy usage. A flow chart and user story for the proposed tool are provided. Direct methods and two-phase methods are considered.

ABSTRACT. Healthcare networks face escalating Advanced Persistent Threat (APT) attacks that spread silently across Internet of Medical Things (IoMT) devices. Existing intrusion detection systems either classify individual flows in isolation or represent multi-device interactions as pairwise edges, discarding the group-level causal structure essential for detecting coordinated attacks. Both categories also absorb clinical traffic confounders such as firmware update bursts and device-polling cycles as background noise, inflating false positive rates (FPR) during routine operations. We propose a four-module pipeline that addresses these limitations jointly. HPCL constructs typed clinical hyperedges from multi-source audit logs, encodes them via a variational autoencoder, suppresses spurious correlations through Pearl’s backdoor-adjusted causal attention, and enforces a formally certified FPR bound using isotonic-regression-calibrated counterfactual pruning. To our knowledge, HPCL is the first IoMT APT detector to simultaneously resolve the pairwise-edge bottleneck, eliminate healthcare-specific confounders, and provide a clinically certifiable FPR guarantee.