ABSTRACT. Interactive session Facilitators: Michaela Emch, Daniel Ari Friedman, Shereen Amos

ABSTRACT. Facilitators: Michaela Emch, Daniel Ari Friedman, Shereen Amos

ABSTRACT. Reconstructing the dynamics of complex networks from observational data remains a central challenge in fields ranging from neuroscience to power grid engineering. While recent approaches combining model reduction with sparse recovery have shown promise for weakly coupled chaotic systems on scale-free topologies, they rely on assumptions about coupling structure and require explicit basis function libraries. In this work, we propose a reservoir computing framework for reconstructing network dynamics of coupled discrete chaotic units directly from time series data. Our approach leverages the inherent capacity of echo state networks to approximate nonlinear dynamical systems without prior knowledge of the governing equations or coupling functions. We demonstrate that reservoir computing can simultaneously infer the local node dynamics, the interaction functions, and the connectivity structure from observations. We benchmark our method against sparse regression-based techniques on both synthetic scale-free networks and biologically realistic neuronal topologies, including a mouse neocortex connectivity. Our results show that reservoir computing offers competitive reconstruction accuracy while relaxing key constraints on network topology and coupling regime. Furthermore, the trained reservoir model enables prediction of emergent collective behavior, including critical transitions to synchronization under parameter variation.

ABSTRACT. The main focus of this paper is how spatial variables on city-scale enable the spread of airborne disease (Covid-19). Data from Ireland’s Central Statistics Office was modelled to identify the relationship between road network accessibility and Covid-19 cases. We focused on three cities; Dublin, Cork, Galway. Of interest was to learn where and when Covid-19 cases rise and how that is impacted by spatial access to healthcare facilities. Multivariate analysis of accessibility and density showed a slight correlation between number of Covid-19 cases and centrality closeness of street network. When modelling this data per month in Dublin, we found that the spread of infections is initially triggered by personal and social reasons, but the rise of infections is driven by road network accessibility. We also modelled survey data conducted at University College Cork to understand the relationship between socioeconomics and environmental and social activities pre Covid-19 and during Covid-19 lockdowns in Cork. We found good correlations between, accessibility, number of landuse types visited and number of covid-19 cases. Further research is needed to verify these findings in other cities, exploring in greater details the impact of spatial accessibility on the performance of healthcare facilities in the event of pandemics.

ABSTRACT. Although it is well-established that ants are capable of responding to fluctuations in the quality of foraging resources, relatively few studies have explored how they adapt when a previously available food source becomes inaccessible. To address this gap, the present study investigates how ants adjust their foraging behavior when the location of the food source is periodically altered. We conducted controlled experiments using Lasius niger, a species commonly employed in pheromone communication research. The experimental design involved a colony of garden ants foraging in an open-field setup, consisting of a 50 cm × 50 cm PVC board with a slope connecting the nest to the field. The food source was relocated between two distinct positions every 10 minutes. Two experimental conditions were established, one with a narrow angular separation (30 degrees) between the food positions, and another with a wider separation (60 degrees). For comparison, a control condition was implemented in which food was simultaneously available at both positions throughout the trial. The results indicated that ants in the control condition consistently exhibited a preference for one of the food sources. In contrast, when the food source was moved in the experimental condition (the main experiment), the ants demonstrated a flexible adjustment in their foraging preference, successfully tracking the location of the food source (Figure 1). Moreover, the distance between the two food positions was found to significantly influence the ants’ foraging motivation. In conclusion, the findings provide evidence that ants possess the ability to adapt to the rearrangement of foraging resources, adjusting their preferences and behaviors accordingly.

ABSTRACT. Frailty syndrome in older adults is a multifactorial decay in reserve and function across several organ systems, in which a minor illness (e.g., influenza) can cause disproportionally large declines in health; yet traditional approaches assess it through isolated clinical deficits and fail to capture whole-system behavior. Foundational network models [1][2] established damage propagation dynamics in large-scale networks, but whether the same emergent and self-organizing properties hold in a minimal network built solely on core phenotype criteria remains unexplored. This study implements a 4-node stochastic network-based model (Figure 1) using frailty phenotype described in [3], where individuals are classified as robust, pre-frail, or frail according to their accumulated damage proportion. A synthetic dataset of 200 subjects was generated from thresholds defined in [3], acknowledging that this may not fully capture real clinical variability. Simulation experiments show that network coupling is the necessary condition for frailty to emerge – when coupling is removed, damage remains contained below the frailty threshold regardless of initial state (Figure 2). Further, a single damaged node is sufficient to cascade the system past the frailty threshold, with a nonlinear tipping point at two simultaneously compromised nodes. The system self-organizes into three discrete stable states driven solely by internal dynamics, with repair resistance serving as the bifurcation parameter. These findings introduce a complexity-based understanding of frailty progression with implications for early detection and intervention in aging populations.

ABSTRACT. Recovering latent structure from count data has received considerable attention in network inference, particularly when one seeks to estimate both cross-group interactions and within-group similarity patterns in bipartite ecological networks. Such networks are often sparse and imperfectly detected, so the observed interaction patterns are biased realizations of the underlying ecological structure, motivating the recovery of latent relationships. Existing models such as Poisson N-mixture models primarily focus on interaction recovery, while the induced similarity graphs are much less studied. Moreover, sparsity is often not explicitly controlled, and scale imbalance among latent factors can lead to oversparse or poorly rescaled estimates with degrading structural recovery. To address these issues, we propose a framework for structured sparse nonnegative low-rank factorization with detection probability estimation. While retaining the two-layer structure of Poisson N-mixture models, we impose nonconvex ℓ1/2 regularization on the latent similarity and connectivity structures to promote sparsity in within-group similarity and cross-group connectivity while preserving relative scale. The resulting optimization problem is nonconvex and nonsmooth. To solve it, we develop an ADMM-based algorithm with adaptive penalization and scale-aware initialization, and establish its asymptotic feasibility and KKT stationarity of cluster points under mild regularity conditions. Experiments on synthetic and real-world ecological datasets demonstrate improved recovery of latent factors and similarity/connectivity structures relative to existing baselines, as shown in Figure 1. Sparsity constraints effectively remove weak connections, yielding interaction patterns consistent with known ecological structures. These results show that integrating Poisson count modeling, detection probability estimation, and sparse low-rank factorization provides a robust framework for inferring ecological networks from environmental sensor data characterized by missing, noisy, or biased observations.

ABSTRACT. Drug-Drug interactions pose a major challenge in modern healthcare and their complexity has increased even more because of polypharmacy. Interactions between different drugs might cause adverse physiological effects or significantly reduce the effectiveness of either drug. Testing for DDIs using lab tests or clinical trials is inefficient and costly, as they can neither study the large combinatorial possibilities of drug pairs nor keep up with the rapid rate of new drugs. Therefore, computational approaches have emerged as a scalable and efficient alternative for predicting potential interactions [1,2]. In this study, a machine learning-based framework is used to predict drug-drug interactions using molecular information from the chemical structure of drugs. The system is trained on the DrugBank DDI dataset with 191,808 drug pairs comprising 1,706 different drugs and 86 interaction types. Drugs are represented by their SMILES (Simplified Molecular Input Line Entry System) string and then converted to a molecular fingerprint using cheminformatics methods to form the features that would be input into a machine learning model [3]. These fingerprints are combined to form a high-dimensional feature vector representing the drug pair. Random Forest and XGBoost were used as ensemble models for multi-class classification of drug interaction types. Experimental results show that Random Forest has the highest accuracy of 86.8% while XGBoost has an accuracy of 83.6%. A strong prediction was demonstrated by the Top-k accuracy: Top-3 was 97.75% and Top-5 was 99.52%, indicating the effectiveness of ensemble learning in capturing complex nonlinear relationships between molecular features. The proposed framework demonstrates that enough information can be extracted from the molecular fingerprints derived from the chemical structure to effectively predict drug-drug interactions. Compared to emerging deep learning approaches such as graph neural networks, the presented method offers a computationally efficient and interpretable alternative while maintaining competitive performance [2,4].

Figure 1. Drug- Drug interaction Workflow Using Molecular Feature-Based Machine Learning

ABSTRACT. During the October 2024 DANA flood in Valencia, spontaneous volunteer groups on Telegram became crucial hubs for coordinating aid. Nonetheless, these self-organized communities were also exposed to disinformation and polarizing content, which could potentially disrupt their collective function. For that reason, understanding the fundamental temporal patterns of communication in such groups is essential for assessing their resilience and for building models that can capture the impact of information disorders. We analyzed message data from the “Voluntarios Dana Valencia” Telegram group during the two months following the flood (until January 1, 2025). Using methods from statistical physics, we characterized user activity by fitting waiting-time distributions to a truncated power law and by applying the time-rescaling theorem to evaluate the consistency of the data with a rescaled Poisson process[1]. We also computed the Fano factor to quantify clustering of message events across time scales. Furthermore, we also compared waiting-time distributions for messages originating from untrustworthy sources and from users identified as high-frequency spreaders of disinformation to assess their influence on the collective communication dynamics. The waiting-time distribution for the community and for individual channels with over 1000 messages follow a truncated power law with exponents α between 3/2 and 5/2, placing the dynamics in the “highly attentive regime” of human communication[2]. The time-rescaling transformation produced rescaled intervals that closely match a unit-rate exponential distribution, with only minor deviations at very short times. The Fano factor is consistently greater than one across all bin sizes, confirming significant over-dispersion and bursty activity. The observed temporal patterns align with known universality classes of human dynamics, while the success of the time-rescaling theorem provides a strong quantitative baseline for modeling. The small discrepancies at short, rescaled times and the persistent over-dispersion suggest subtle, non-Poissonian features that may be linked to rapid response cascades or the injection of disinformation. This work establishes a foundation for the DRI (Desinformació, Resiliència i Infraestructures) project’s broader goal of quantifying how information disorders affect post-crisis self-organization.

ABSTRACT. In this work, we propose a first-principles meta-model of the behavioral sink, a form of demographic collapse first observed by John B. Calhoun in the Universe 25 experiment, in which a confined population collapses despite the absence of predation, disease, or immediate resource scarcity. The meta-model interprets this dynamic as an endogenous collapse driven by overcrowding, understood as perceived density rather than as a purely physical measure of population density. Starting from this conceptual framework, we develop two representations of the phenomenon, an EBM and an ABM. The two models share the same structure and parameterization, enabling direct comparison across representation styles while minimizing the influence of modeling assumptions related to the unit of analysis. The parameters are calibrated on the EBM and then applied unchanged to the ABM. Results show that the EBM reproduces the empirical growth–collapse trajectory with low error, while the ABM generates a structured set of possible trajectories organized into a limited number of recurrent regimes. The dominant regime is consistent with the empirical data and remains stable under parameter perturbations, indicating that the observed dynamics correspond to the most probable system behavior rather than to a specific realization. These findings suggest that a minimal meta-model based on density-dependent stress feedback is sufficient to explain endogenous collapse dynamics and may provide a general framework for analyzing similar processes in ecological and socio-ecological systems.

ABSTRACT. Major Evolutionary Transitions (METs), in which previously independent entities integrate into higher-level individuals with emergent capabilities, represent the most consequential events in the history of life. The existing literature describes what transitions look like (Maynard Smith and Szathmary, 1995) and the fitness dynamics that accompany them (Michod, 1999), but does not identify a general causal mechanism that generates them across all substrates and scales. We propose that METs are driven by two coupled feedback loops operating under substrate-independent physical constraints.

The first loop, the Vulnerability-Cooperation Paradox, explains why entities become integrated: cooperation produces specialization, specialization increases vulnerability, vulnerability demands deeper cooperation. The ratchet is irreversible, operating from gene consolidation through endosymbiosis, multicellularity, colonial integration, and human civilization. The Black Queen Hypothesis (Morris et al., 2012) demonstrates that this outcome arises even through purely selfish dynamics.

The second loop, the Scale-Communication Challenge, explains whether integration succeeds or fails: growth overwhelms existing communication systems, forcing qualitative leaps to new modalities or producing fragmentation. We derive an 8-stage communication sequence and use Argentine ant supercolonies as a comparative case: continental-scale colonies that persist genetically but lose functional coordination because chemical communication cannot sustain integration at that scale (Moffett, 2012).

We formalize the coupling as ODEs governing specialization (S), scale (N), and communication (C) with a scaling exponent (delta) as control parameter. METs correspond to critical transitions where both loops resolve simultaneously. The model generates 10 testable predictions distinguishing it from existing MET frameworks, including predictions about the current human-technology integration and the role of LLMs as structured aggregation of collective human cognition (Abbas, 2009; Siththaranjan et al., 2024; Vojnovic and Yun, 2025). Full paper: DOI 10.5281/zenodo.19447319.

ABSTRACT. Large language models in advisory and decision-support roles are expected to hold and defend positions. When users push back repeatedly, the question is not only whether a model eventually gives in, but how it behaves across a prolonged disagreement: whether it repeats itself, becomes more assertive, or shifts in emotional tone. We present a pilot analysis of four LLMs (Claude-Sonnet, Gemini-2.5-Flash, GPT-5.4, Llama-4-Maverick), each assigned a position on six debate topics and exposed to 60 turns of pushback across four conditions: simple rejection, rejection with counterargument, neutral acknowledgement, and taunting. Across all models and resistance conditions, no explicit capitulation was observed. Lexical diversity remained stable, but arguments became less novel over time: later responses were more similar to at least one prior response than early ones, across all four models. Models differed in how they constructed these arguments: Claude used roughly four to six times more named entities per 100 words than the others, consistently anchoring responses in specific facts. relatively high certainty marker frequency, though both certainty and hedging declined gradually over turns. GPT and Gemini were comparatively stable, though hedging drifted downward in both. These patterns suggest that sustained resistance does not leave a model's argumentation unchanged, even when its stated position holds: each model shifted its rhetorical style in ways that were consistent and model-specific. The most striking emotional pattern came from GPT: as conversations progressed, both anger and trust language rose together, suggesting growing intensity alongside sustained rapport-seeking. Gemini and Llama showed the opposite on anger, which generally fell over turns, though trust was less consistent across the two. Claude stood apart: both anger and trust declined, consistent with a more fact-anchored, less emotionally engaged style. Tracking position change alone misses most of what happens when users push back: models held their ground but shifted how they argued and how emotionally they engaged, in ways that differed substantially across models. A user pushing back against GPT faces growing emotional intensity alongside sustained rapport-seeking; one pushing back against Claude faces a model that is more fact-anchored and less emotional. These are meaningfully different persuasive encounters, invisible to compliance-only evaluation. This study is limited in scale and uses simulated resistance; replication with real users is needed.

ABSTRACT. Most models of customer behavior assume customers act as rational utility maximizers. Behavioral economics shows that real customers often do not, especially under financial pressure or after repeated commitments to a brand. This study tests two bounded rationality mechanisms so that future agent-based models can rest on empirical evidence rather than the rational-actor assumption. Using the Customer Personality Analysis dataset (N = 2,206 after cleaning), we test two hypotheses. H1 holds that customers under economic pressure show a higher present bias ratio - the weight placed on immediate deal-based rewards relative to longer-horizon goods. H2 holds that customers who accepted a past campaign are more likely to accept the current one. A Mann-Whitney U test shows pressured customers have a median present bias ratio 2.64 times higher than other customers (p < .001), stable across income thresholds and winsorization. A logistic regression controlling for behavioral state and recency shows past acceptors have about seven times the odds of accepting the current campaign (odds ratio = 6.95, p < .001), with a clear dose-response from 8% to 91%. Both hypotheses are supported. Figure 1 summarizes these findings as a conceptual research model. Both hypotheses are supported, so this research contributes two new bounded-rationality attributes to the agent design: an economic-pressure attribute and a commitment attribute. Prior research [1, 2, 3] supplies the other agent-based model (ABM) features network topology, cascade propagation, and behavioral states. Together, these features set up the next research question: does adding bounded rationality to the state-aware policy improve cascade containment?

ABSTRACT. The global Bahá'í community has developed an evolving "framework for action": interlocking practices, institutional capacities, and learning processes spanning community development, social action, and engagement with the discourses of society. The features of this framework mirror many of the features of complex human systems, including mutual dependencies, feedbacks, delays, and multiple levels of scale. Systems science at Association for Bahá'í Studies (ABS) traces to a 1986 Ervin Laszlo address on general systems theory, cybernetics, and far-from-equilibrium thermodynamics in relation to Bahá'í teachings on unity and evolution[1]. The working group resurfaces this thread with contemporary methods, most notably group model building[2]. Formed after a 2024 ABS presentation by L. Kurt Kreuger[3], in 2025 we facilitated a participatory seminar on causal loop modeling of the framework[4]. Online meetings have since pursued causal loop diagrams, system dynamics and agent-based models, examining feedback structures, dynamics, and emergent conditions for collective capacity. Loops linking social, devotional, and educational activities recur as engines of nonlinear growth, and participatory modeling itself surfaces assumptions implicit in prose descriptions. We invite collaboration with researchers of religion, learning communities, and large-scale collective action.

ABSTRACT. Criticality in infrastructure systems is often quantified using indicators that capture different aspects of node importance, including structural position, operational load, and cascading failure potential. However, these measures often produce inconsistent rankings, making it difficult to determine which components should be prioritized for system-level intervention.

This inconsistency reflects a deeper question: how local magnitude-based indicators (e.g., nodal load) relate to system-level importance arising from network interactions. While complex systems research emphasizes that criticality emerges from system-wide dependencies, engineering studies often associate high-load nodes with increased operational importance. It remains unclear when local indicators can reliably approximate system-level importance, and when they fail due to system-wide interactions.

This study investigates this relationship using a proxy evaluation framework. We model infrastructure as a state-dependent functional network and use eigenvector centrality as a proxy for system-level importance, while treating nodal load as an observable local indicator. Using nodal time series data from the ISO New England power system, we evaluate the alignment between these measures across baseline and peak operating conditions, where peak is defined as the top 5% of system load.

It is found that load-based rankings exhibit moderate alignment with network-derived centrality under baseline conditions, but this alignment degrades significantly under peak demand. Further analysis shows that this breakdown is not directly driven by load magnitude, but by increased concentration of system-wide coupling. A minimal generative model shows that as coupling strength increases, alignment between local indicators and system-level importance rapidly collapses.

These results suggest that different dimensions of criticality are not universally aligned, and that their relationship depends on the level of system coupling. This work provides a framework for evaluating when observable indicators can reliably approximate system-level importance in complex infrastructure systems.

ABSTRACT. One way to assess the importance of a node in a network is to observe what happens to the size of the largest connected component ($LCC$) after that node is removed. If we keep removing nodes we obtain a disconnected network [1]. For each node $i$ we have $N-1$ other nodes reachable; we pick the paths that minimize $K_{ij} = \min(\prod_{h=i}^{j-1} k_{h})$, where $j$ is the target node and $k_h$ is the degree of node $h \in \mathcal{P}(i,j)$. Among those paths we pick the shortest path. We define the normalized generalized Shannon entropy ($NGSE$) as: $ S_{i}^N = \sum_{h \neq i}^{N} \frac{\log_2 (Q_{i}{K}_{ih})}{{Q_{i}K}_{ih}}$ where $Q_i = \sum_{h \neq i}^{N} \frac{1}{K_{ih}}$ is the normalization factor.

$NGSE$ has been used to perform targeted attacks on Watts-Strogatz ($WS$) networks [1] alongside classical centralities such as Degree, Betweenness, Closeness, and Eigenvector centrality; random node removal serves as a benchmark. The more significant a node, the smaller the fraction of removed nodes ($f_r$) required to reduce $LCC$ to chosen thresholds of $10\%$ and $5\%$. We generated $100$ $WS$ networks with $N=400$, $k=8$, and $p=0.04$ and report the average $LCC$ as a function of $f_r$ for each strategy. As seen in Figure~1 (left), $NGSE$ identifies key nodes early, reaching the $10\%$ $LCC$ threshold after removing $54.4\%$ of nodes; for the $5\%$ threshold the best strategy is Betweenness.

As a real-world case study, we analyzed the Great Britain power-grid network ($GBPN$) [2], composed of high-voltage nodes connected by links. We considered only the giant connected component, ignoring isolated nodes, and treated the network as undirected and unweighted. Since the power-grid is uniquely defined, only the random strategy was iterated $100$ times. The connected grid comprises $494$ nodes and $598$ links. As shown in Figure~1 (right), Degree, Betweenness, and $NGSE$ are all effective at disrupting the network; however, $NGSE$ reached the $10\%$ threshold after removing only $17.4\%$ of nodes.

It has been shown that $NGSE$ can identify the most important nodes in $WS$ networks and perform targeted attacks that outperform classical centralities in the early phases of disruption. Applied to the real-world $GBPN$, this measure reveals network weak points that other centralities would ignore, potentially helping to mitigate risks by protecting nodes that appear unimportant by conventional metrics.

ABSTRACT. This paper investigates the emergent properties of generative AI systems, framing these systems as a distinct medium of procedural translation that reconfigures the feedback loops between human intent and algorithmic execution. The study moves beyond the view of AI as a simple tool; instead, it argues that AI-driven creativity is a co-constructive process where agency is distributed among human and non-human actors within a coupled system. Based on Lev Manovich’s "software as medium" and N. Katherine Hayles’s "technogenesis," the research traces the trajectory of an idea from the initial human linguistic sparks to their translation into autonomous algorithmic flows. Central to this inquiry are the systemic constraints of translation, which represent the moments where human intent dissolves into machine-to-machine (M2M) communication. Through an analysis of the interactive work M to M: Self-Organizing AI Dialogues [2025] and practice-based research in non-linear data mapping, such as CO2 Emissions, this research uncovers how meaning emerges from the entanglement of algorithmic logic and human curatorial intervention. Ultimately, it contends that AI art represents a radical procedural translation of human consciousness; this process demands a new ethical and aesthetic framework for our technological co-evolution.

ABSTRACT. “Are you Christian?” This was the very first question Ellie, an 86-year-old white woman in a nursing home- who would pass 6 weeks later- asked me. Being born in 1985, in communist Russia, growing up in Germany, and socialized into Christian civic religion, this question seemed rather odd to me. However, the profound interaction with Ellie inspired me to explore the multifaceted evolution of religious identity across the life course and to frame it as a dynamic and complex social process. Drawing on over 30 hours of in-depth qualitative interviews with three elderly women—Ellie (86), Molly (80), and Terry (78)— this paper analyzes how and why an individual´s religious self is shaped, maintained, and transformed over time. The findings show that religious identity is highly contextual, malleable, and continually reconciled within complex networks of socio-cultural environments, interdependent "linked lives," key life transitions, and broader societal cohort replacements. To conceptualize these continuous shifts, I eventually introduce "The Arch of Religious Identity", a framework outlining three primary functions of faith that drive identity adaptation across different life stages. First, the overarching process is lifelong Socialization. Foundations of early religious identity can be established through caregiving environments that build basic trust and embed psychological "ontological security" via religious social behavior. Second, as individuals transition into the complex social spheres of adulthood, religious identity shifts to serve an Instrumental Utility. During this phase, it becomes highly flexible and draws on external religious structures to maintain family stability and social cohesion. Finally, in late life, the religious self undergoes a profound transformation as the function of faith shifts toward securing Existential Meaning. In line with the psychosocial theory of Gerotranscendence, external religious roles and dogmas often fall away. If replaced by an intrinsic and personal spiritual integration, a transcended identity enables individuals to face mortality without overwhelming anxiety. The research highlights how religious identity within shifting social structures transitions from externally reliant childhood trust to late-life transcendent wisdom.

ABSTRACT. Patient-specific blood flow simulation is a powerful tool for estimating vascular properties and hemodynamic conditions that are difficult to measure directly. However, solving the corresponding inverse problem is challenging due to its ill-posed nature as well as the difficulty of differentiating the forward process. In this work, we propose a Gaussian process-based Bayesian optimization framework for parameter estimation in one-dimensional blood flow simulations using openBF. Since openBF is not differentiable, gradient-based optimization methods cannot be directly applied. Instead, Bayesian optimization treats the simulator as a black-box model and builds a probabilistic surrogate of the objective function from a limited number of forward simulations. The objective function measures the mismatch between simulated and observed hemodynamic signals, such as pressure and flow rate, at selected vascular locations. A Gaussian process surrogate is used to estimate both the predicted loss and the uncertainty over the parameter space, while an acquisition function guides the selection of new simulation points. This strategy enables efficient exploration and exploitation, reducing the number of expensive forward simulations required to identify optimal parameters. The proposed framework provides a practical and derivative-free approach for calibrating blood flow models and can support patient-specific cardiovascular modeling when direct parameter measurements are unavailable.

ABSTRACT. In this work, we develop and analyze a mathematical model of pre- frontal–amygdala circuitry using Wilson–Cowan population equations to capture the excitatory–inhibitory dynamics underlying emotional control. The model incorporates reciprocal projections between prefrontal pyramidal cells and basolateral amygdala principal neurons, as well as inhibitory microcircuits that me- diate feed-forward and feedback suppression. To represent the influence of stress hormones, we augment this neural framework with a cortisol-dependent delay term. Specifically, we implement an integral kernel that encodes the system’s memory of past arousal activity: acute stress is modeled as a sharp kernel emphasizing recent amygdala firing, whereas chronic stress corresponds to a broader kernel integrating over a longer history. Through numerical simulations and bifurcation analysis, we characterize the system’s behavior under baseline, acute, and chronic stress conditions. We focus on the emergence of distinct amygdala firing rhythms, their modulation by prefrontal feedback, and the phase transitions between dynamical regimes.

ABSTRACT. Immigration produces diverse outcomes, with some immigrants finding happiness in an improved quality of life, while others face emotional distress from unexpected challenges. Studying these emotional experiences is challenging due to limited longitudinal data. To address this, we curated high-quality data and analyzed Twitter activities of immigrants in the United States to explore how their expressed sentiments evolve post-migration, comparing them to non-immigrants and US locals in a quasi-causal sense. Our findings revealed that migration generally boosts positive sentiments among immigrants, even when discussing the same topics, and that this effect lasts up to one-year post-migration. However, there is no conclusive evidence that migration significantly impacts negative sentiments. These patterns indicate that the migration event is associated with an expansion of positive expressed sentiments among immigrants, rather than a reduction in negative ones, suggesting a reorientation of emotional expression that reflects adaptation or opportunity recognition following migration. Additionally, age and gender play a key role, with younger population and females more likely to express positive sentiments post-migration, while linguistic ties also help foster increased positivity. These insights advance our understanding of immigrants’ emotional adaptation and highlight key demographic and social factors that shape their post-migration experiences.

ABSTRACT. Clinical language model systems that operate on report text are increasingly deployed as multi-stage pipelines in which retrieved or constructed evidence is passed to downstream decision components. However, component-level performance does not reveal how upstream errors affect downstream predictions. We study failure propagation in a controlled two-stage clinical language pipeline using the Findings section of radiology reports from MIMIC-CXR. We introduce systematic perturbations to negation and uncertainty expressions in upstream findings text and evaluate paired downstream predictions under true and corrupted evidence using Qwen/Qwen2.5-7B-Instruct and flan-t5-base. Across 3,111 paired cases, moderate corruption produces measurable prediction changes (7.1%) but only a small and statistically uncertain decrease in aggregate accuracy, indicating bounded propagation under limited evidence degradation. However, severity analysis reveals a non-linear response: strong corruption causes substantial downstream failure, with large increases in harmful prediction transitions. We further observe that propagation behavior is model-dependent and varies across diagnostic categories. These results suggest that clinical language pipelines can remain relatively stable under moderate perturbations but fail sharply once upstream evidence distortion exceeds a severity threshold. Paired pipeline-level evaluation therefore provides a practical framework for identifying harmful transitions and robustness limits that are not captured by component metrics or aggregate accuracy alone.

ABSTRACT. AI research on multi-agent systems and management science both study coordination, decision-making, and team dynamics, yet these fields have developed largely in isolation. Decades of management research have produced well-validated principles for human teams that may offer valuable knowledge transfer opportunities for designing AI multi-agent systems. To explore this potential, we map the conceptual landscapes of both fields to identify where they converge and diverge. We analyze 88{,}048 management papers from expert-selected journals and 6{,}512 AI papers from arXiv filtered by keywords such as ``multi-agent systems,'' ``large language models,'' and ``LLM agents,'' embedding each into a 256-dimensional vector. We extract domain-specific keywords (238 management, 165 AI) via TF-IDF with LLM-based filtering. Each keyword is contextualized by embedding it within its surrounding sentences and retrieving the token-level vector, capturing domain-specific usage rather than surface-level semantics. Agglomerative clustering of cross-domain keyword similarities reveals no large block of strongly associated shared keywords across both domains, direct evidence of discommunication between the fields. A small number of keywords do appear in both domains (e.g., ``strategic decision-making,'' ``problem solving''). Using them as anchors, Vec2Vec aligns both embedding spaces into a shared representation. The aligned space (Fig.~\ref{fig:aligned}) reveals management-dominated regions (blue) around organizational commitment, job satisfaction, and HR management; AI-dominated regions (red) around LLM-based agents, agentic workflows, and AI safety; and shared regions (white) where decision-making, problem solving, and social simulation converge with individual differences and job performance, highlighting concrete opportunities for cross-domain transfer.

ABSTRACT. In traditional cultural transmission models, the interaction network is typically generated first, followed by the application of dynamics over it. However, in real-world societies, the formation of social networks and the dynamics of opinion or cultural formation occur simultaneously.

In this communication we present the integration of Axelrod dynamics on continuously growing networks. In our model, each incoming node establishes m initial connections with existing nodes using a uniform attachment probability. The Axelrod model defines each individual as an F-dimensional vector, where the number of components indicates distinct cultural features, and q represents the cultural variability. Agents interact with a probability matching their cultural overlap, assimilating differing traits after contact.

Using a master equation approach, we analytically derive the properties of the network. As shown in Figure 1, the degree distribution follows the relation P(k) = [1 / (m+1)] * [m / (m+1)]^(k-m). Additionally, in Figure 2, we show the expected degree of nodes k_s(t) as a function of their birth index s at a fixed simulation time t, governed by k_s(t) = m + m * ln(t/s). To approximate real social behavior, we incorporate a "friendship network" mechanism driven by triadic closure, demonstrating that these growing networks fulfill Small-World properties. Finally, during the presentation, we will detail the results of integrating Axelrod dynamics on growing networks, showing how continuous growth disrupts global consensus and shifts the system toward cultural fragmentation.

ABSTRACT. On election day, hundreds of millions of people, each carrying their own mix of values, grievances, half-remembered news stories, and conversations with friends, walk into a booth and make a choice. That chaotic swirl of individual decisions becomes a government. Political science has spent decades studying voters one at a time: their beliefs, their biases, their likelihood of turning out. We think that lens, while valuable, misses something important. The outcome of a democratic election is not just the sum of individual choices. It is what emerges when those choices interact across a particular social structure. That is a complex systems problem, and we treat it as one. We model the democratic polity as a network. Voters are nodes, and the edges between them represent who talks to whom, who influences whom, who shares the same information environment. Drawing on bounded-confidence models of opinion dynamics, we find that what determines whether a society converges toward shared understanding or fractures into warring camps is less about what individuals believe and more about the shape of the network they inhabit. When that network is well-connected, when there are many paths between different kinds of people, opinions tend to shift, moderate, and eventually find some common ground. When the network is modular, carved into dense clusters with few bridges between them, something different happens. People do not become more extreme because they are irrational. They become more extreme because every conversation they have confirms what they already think. The network itself is doing the pushing. Our central claim is that healthy democracies operate near a critical point, connected enough to absorb new information and update, coherent enough to reach decisions that stick. We look at three reform proposals that complexity theory suggests might actually move the needle. Deliberative democracy, Sortition and Structural intervention. A way of rewiring the network so that the system's own dynamics pull it back toward the critical regime, where real collective intelligence becomes possible again.

ABSTRACT. International Economic Integration (IEI) is a multidimensional process that traditional trade theories often fail to capture due to their linear nature. This work presents a robust integrated framework that applies Economic Complexity (EC) theory to regional blocs treated as unified entities. By aggregating national export matrices, we simulate the "Combined Productive Capacity" of international alliances such as the European Union, MERCOSUR, the Andean Community, and URUPABOL. The framework acknowledges the fundamental contributions of both the Economic Complexity algorithms —Economic Complexity Index (ECI) and Product Complexity Index (PCI)— and the Economic Fitness algorithms —Countries Fitness Index (Fc) and Product Complexity Index (Qp)—, utilizing their complementary strengths to quantify collective capabilities. The methodology is supported by a systematic literature review (2020-2025) of IEI, addressing the identified research gap regarding quantitative tools for evaluating the dynamic effects of institutionalized integration. Experimental validation through MATLAB® simulations highlights the model's superior ability to capture synergies that exceed individual national potential. By treating regional blocs as unified entities, the framework effectively quantifies how integration serves as a powerful mechanism for collective capacity absorption, transforming product diversity into a measurable competitive advantage. Furthermore, the model proves its strategic value by establishing a direct link between integrated industrial complexity and the Human Development Index (HDI), proving that the accumulation of collective capabilities is a fundamental driver for long-term social welfare. This framework provides policymakers with a novel, high-impact computational tool to design and evaluate international economic strategies in an increasingly fragmented global landscape.

ABSTRACT. The Deffuant bounded-confidence model is an agent-based model in which agents hold continuous opinions in [0, 1], interact pairwise, and move toward each other only when their opinion difference falls below a confidence threshold ε. The number and arrangement of opinion clusters that emerge depend on ε, the network through which agents interact, and the size of the population. Cluster counts also depend on a tolerance parameter, the threshold below which two distinct final opinions are counted as belonging to the same cluster. To understand the dynamics, we have to separate its effects from those of the tolerance.

We ran the Deffuant model on a complete graph (N = 100 agents, convergence parameter μ = 0.5, T = 10,000 update steps) at ten values of the confidence threshold ε in [0.02, 0.20], and at each value counted clusters using ten different tolerances from 0.001 to 0.05. Clusters are defined by sorting the final opinions and grouping any consecutive pair whose gap falls below the tolerance.

The relationship between cluster count and tolerance changes shape qualitatively between ε = 0.02 and ε = 0.04. At ε = 0.02 the cluster count we get varies by a factor of ten across this tolerance range, from 33.4 ± 2.9 at tol = 0.001 to 3.6 ± 1.2 at tol = 0.05. At ε = 0.04 this spread is already much smaller, and by ε ≥ 0.06 the cluster count barely depends on tolerance, which tells us that at higher confidence thresholds the final clusters are separated by gaps larger than 0.05 in opinion space.

To verify that this pattern reflects the dynamics and not merely the cluster-counting procedure, we computed a null baseline by counting clusters in 1,000 draws of N = 100 uniform random opinions in [0, 1] at each tolerance. The post-dynamics cluster counts at ε = 0.02 fall well below the null at fine tolerance (33.4 vs. 90.5 ± 2.9 at tol = 0.001), confirming substantive chained merging: the dynamics has collapsed roughly 57 random clusters into chained groups that remain close together in opinion space. At coarser tolerance the post-dynamics counts exceed the null (3.6 vs. 1.6 ± 0.7 at tol = 0.05). Together, these results indicate a shift in the dynamics of chained merging as the source of the qualitative shift, and motivate further investigation of this phenomenon.

ABSTRACT. Meanings around climate change are not straightforward translations of climate science but rather emerge from complex networks of meaning-making involving political, economic, pop culture, commercial, and religious influences. Traditional climate communication emphasizing catastrophic effects often backfires, leading to helplessness, alienation, and denial rather than action. Alternative approaches---framing climate change as addressable or using artistic representations---show greater promise. However, effective local communication requires understanding local perceptions and emotions, which remain poorly understood. This study, connected with the Binghamton 2 Degrees community resilience initiative, seeks to build theory about why some individuals see climate change as present and pressing while others do not.

We present a novel approach: constructing AI ``digital twins'' from 91 semi-structured interviews with community members in Binghamton, NY. These interactive avatars, built using large language models (LLMs) trained on interview transcripts capturing demographics, climate stories, emotions, beliefs, and information networks, allow stakeholders to dynamically explore local climate perspectives. We construct digital twins representing distinct community segments---climate-concerned versus climate-skeptical, urban versus rural, different generations---enabling simulated conversations within groups to understand internal discourse and between groups to explore how disparate perspectives engage. The first specific deliverable is simulated comments and focus groups about Broome County's current Climate Resiliency Plan, which is in public comment period. Ideally, these simulated comments from Binghamton residents' climate stories can help the local stakeholders designing this plan.

ABSTRACT. All attendees will be involved. The main event. Participants build out what a complex systems curriculum across SUNY could actually be: core concepts, course sequences, and what each campus is positioned to contribute.