ABSTRACT. Careers are often defined by periods of concentrated achievement, when high-impact work occurs in close succession. Yet behind every celebrated first act lies a deeper question: what happens next? Understanding how a second act unfolds is essential to explaining the long-term dynamics of success. Competing theories offer divergent predictions. The Matthew Effect emphasizes “success-breeds-success” dynamics. Yet the same success also reshapes identity, expectations, and networks in ways that narrow subsequent search. These competing forces may operate at different levels, but their net effect is unified: success re-parameterizes the explore–exploit problem. It tilts the payoff surface toward exploitation, raises the perceived and realized costs of pivoting, and constrains feasible exploration through reputational and structural path dependence. Here, we reconstruct the career histories of 21,470 individual scientists, 3,046 film directors, 2,243 artists, and 4,278 authors. To robustly identify high-impact periods within individual careers, we use the hot-streak framework. This allows us to distinguish individuals who achieve repeated success from those who have a single high-impact period. Across all four domains, we find a consistent pattern: the probability of a second act decreases as the impact of the first act increases. Mechanistically, early success is associated with a lock-in effect: the greater the impact of the first act, the more an individual will continue to pursue the same direction. This lock-in effect, in turn, limits exploration and reduces the likelihood of another period of highly successful work. Crucially, lock-in is not inevitable. Mid-career exploration significantly increases the likelihood of experiencing a second act. This exploration, however, is costly. The larger the pivot, the longer the interval between the two acts. Interestingly, the seeds of a second act, when different from the first, are planted in the early stages of the career. Even when pivoting, individuals do not abandon their initial main topic of interest but rather embrace ambidexterity and a balanced portfolio approach. Our findings suggest that the traditional emphasis on early-career promise may be too narrow. In many domains, the age of peak achievement is rising. This reflects a growing burden of knowledge and the lengthening learning curve required to reach the frontier. As a result, the widespread emphasis on early accomplishments may not reliably predict long-term impact. Our novel perspective is potentially informative for funding agencies, investors, and organizations that need to identify and support long, impactful careers, as well as to support more forward-looking hiring strategies. The capacity to continuously recognize opportunities for exploitation, remain open to novelty, and adapt over time may be as important as early signals of potential.

ABSTRACT. Network medicine has become an effective tool for disease therapeutics by modeling human biology as a complex system through the human interactome, the network of all protein interactions in the human body. We found in a recent study that genes associated with the hallmarks of aging, biological processes that drive loss of function and vulnerability to disease, form topologically localized network neighborhoods called ‘aging hallmark modules’ [1], analogous to disease modules. In this work, we develop a systematic pipeline to identify food compounds that potentially slow aging, integrating data from the protein-protein interaction network (PPI), the aging modules, and the protein targets of food compounds. The first step of the pipeline identifies compounds whose targets lie in network proximity to aging genes and can potentially alter the corresponding phenotype. The second step consists of calculating pAGE, a transcription-based metric that uses gene expression data to determine whether a compound reinforces or reverses aging-associated transcriptional patterns. We identify 434 network candidates for promoting longevity, validate our findings by recovering known lifespan-extending polyphenols among top candidates, and propose mechanisms of action for untested food compounds. Our findings illustrate that a network medicine view of human biology uncovers novel molecular mechanisms and provides falsifiable, interpretable hypotheses for experiments on longevity pathways.

[1] Gross, B., Ehlert, J., Gladyshev, V. N., Loscalzo, J., & Barabási, A. L. (2025). Network-driven discovery of repurposable drugs targeting hallmarks of aging. ArXiv, arXiv-2509.03330

ABSTRACT. What occurs in Operating rooms (ORs) is typically analyzed from a procedural perspective, with a focus on surgery phases, technical execution, and workflow optimization. In contrast, the role played by social dynamics in shaping team performance remains overlooked. To address this gap, we propose Time-Expanded Graphs (TEG) in which temporal evolution collapses into a single graph. In TEG, nodes encode team members in a specific time step and typed edges encode temporal and relational dependencies among the team members. We carried out experiments building on an existing dataset of audio-video recordings of simulated surgical procedures, and we enriched it with a set of automatically extracted features and manual annotations. Specifically, we integrated: (i) sixteen node-level behavioral descriptors capturing leadership and teamwork, (ii) five edge-level descriptors modeling interaction quality, and (iii) six graph-level annotations associated with teamwork performance and procedural duration. Moreover, questionnaires to identify counterfactuals were designed to support intervention analysis. Through our approach, we propose a unified learning setting that jointly captures team performance and counterfactual reasoning signals. Features values are aggregated over six minute consecutive windows of recordings. The same time span was used to manually annotate the excerpts. We address both efficiency prediction through TEG Neural Networks and interpretability through a two-layer counterfactual reasoning, represented in Fig. 1: (i) individual-level interventions on node attributes representing interpretable behavioral traits, (ii) interaction-level interventions modifying communication patterns. We explore two complementary strategies for counterfactual generation. The first is a greedy search approach, explicitly evaluated with respect to surgical duration [1] (i.e., improving slow vs. fast procedure predictions). The second is an explanation-driven approach, guided by graph-level information to identify influential structures and interactions. Overall, our work enables both accurate prediction of team performance and prescriptive analysis through multilevel counterfactual reasoning, while consistently outperforming competitive graph-based, temporal, and tempo-relational baselines across efficiency prediction. This bridges graph-based machine learning with social and behavioral sciences in high-stakes collaborative environments.

ABSTRACT. Online forums such as reddit have emerged as places where people can interact with people from a wide variety of backgrounds. In order to effectively engage in discourse, all parties must be receptive to the view points of the others. This paper presents a new categorization of the ways individuals show both receptiveness and dismissiveness to new ideas. Using a set of manually applied tweets it then builds an NLP model to study these patterns at scale. The goal of this paper is to demonstrate the types of topics that generally lead to higher receptiveness or dismissiveness, and to track how the distribution of our labels over time.

ABSTRACT. Network modeling of complex systems has been widely used to uncover important nodes, edges, paths, and modules, typically by leveraging network topology. In weighted graphs, this requires quantifying how strongly nodes interact with each other directly, via edges, and indirectly, via paths. In this framework, a network is represented by a distance (or an isomorphic proximity) graph where smaller (larger) edge weights quantify stronger interactions [\href{https://doi.org/10.1017/nws.2015.11}{1}]. Interestingly, the topology of networks in nature and society contains a vast amount of shortest-path redundancy [\href{https://doi.org/10.1093/comnet/cnab021}{2}, \href{https://doi.org/10.1007/978-3-031-21131-7_11}{3}], where an indirect path is shorter than the direct edge connecting two nodes. This violation of a generalized triangle inequality enables a principled \textit{network sparsification} methodology [\href{https://doi.org/10.1093/comnet/cnab021}{2}], which we implemented in the \href{https://github.com/CASCI-lab/distanceclosure}{\emph{distanceclosure} Python package.}

Given a distance graph $D$ with edge weights $d_{ij} \in [0, \infty)$, its distance closure, $D^{T,g}$, is a graph whose edge weights are the shortest distances computed on $D$ with path length measure $g$. Only the subset of \textit{triangular} edges, which obey the generalized triangle inequality $d_{ij} \leq g(d_{ik}^{T,g}, d_{kj}^{T,g})$, contribute to shortest paths [\href{https://doi.org/10.1017/nws.2015.11}{1}]. These edges observe $d_{ij} = d_{ij}^{T,g}$ and define the \textit{distance backbone} $B^g$ of $D$ for path length measure $g$, which is an invariant subgraph in the computation of $D^{T,g}$ [\href{https://doi.org/10.1093/comnet/cnab021}{2}]. Conversely, \textit{semi-triangular} edges violate the generalized triangular inequality and thus satisfy $d_{ij} > d_{ij}^{T,g}$. Given that they do not contribute to any shortest path (i.e., $D^{T,g}$), removing them reveals the distance backbone $B^g$. Setting $g\equiv +$ yields the \textit{metric backbone} $B^m$ which is typically a small subgraph (average of 70\% edge reduction [\href{https://doi.org/10.1093/comnet/cnab021}{2}]) in natural, social, and technological (directed and undirected) networks, enabling their robustness to attack [\href{https://doi.org/10.1093/comnet/cnab021}{2}, \href{https://doi.org/10.1007/978-3-031-21131-7_11}{3}]. Setting $g\equiv \max$ yields the \textit{ultrametric backbone} $B^u$, which is the union of all minimum spanning forests [\href{https://doi.org/10.1088/2632-072X/ad679e}{4}].

The utility of the \emph{distanceclosure} package has been demonstrated in various studies, revealing that the algebraically-principled distance backbone methodology does not require threshold parameters or null models, and outperforms other sparsification methods in preserving epidemic spreading [\href{https://doi.org/10.1088/2632-072X/adf2ed}{5}] and community structure [\href{https://doi.org/10.7554/eLife.95774}{6}]. The latest version of the package provides four algorithms to compute distance backbones $B^g$: comparison post closure computation (Closure) [\href{https://doi.org/10.1093/comnet/cnab021}{2}], heuristic approximation (Heuristic) [\href{https://doi.org/10.14778/2947618.2947623}{7}], and two iterative exact computations with different memory requirements (Iterative and Flagged) [\href{https://doi.org/10.1007/978-3-031-21131-7_11}{3}]. We will discuss the importance of these differences across networks with varying sparsity and edge weight heterogeneity.

ABSTRACT. Equal occupation does not imply equal collective state. In correlated threshold systems, fields with identical occupied fractions may exhibit fundamentally different connectivity organizations, ranging from fragmented local clusters to dominant-connected and domain-spanning structures. Here, we develop a connectivity-centered framework to characterize this non-uniqueness under strict equal-probability constraints, using a field-derived liquefaction manifestation map—i.e., a spatial record of earthquake-induced ground-failure evidence—as empirical grounding. Coherence-controlled spatial ensembles are generated from a common latent template and thresholded to identical occupation levels. Connectivity is quantified using dominant-cluster ratio, spanning probability, connected-component structure, and mesoscale indicators. Null-model comparisons show that uncorrelated fields remain predominantly fragmented, whereas correlated fields reorganize into clustered, dominant-connected, and spanning regimes under identical occupation constraints. Regime analysis further shows that equal occupation maps not to a unique connected state, but to a coherence-dependent set of accessible configurations. Additional mesoscale and fluctuation analyses suggest that this reorganization proceeds through intermediate transition bands rather than through a single sharp boundary. Together, these findings show that spatial correlation reshapes the accessible connectivity state space of threshold fields and that extent-based interpretations of spatial hazard are fundamentally incomplete.

ABSTRACT. GDP counts value produced locally. It says nothing about which states enable production elsewhere. We construct a weighted, directed network of interstate commodity flows (51 nodes, ~2,500 edges) from the 2017 Commodity Flow Survey and ask a simple question: which states are structurally important to the national trade system in ways GDP cannot see?

Eigenvector centrality provides a clear answer. The GDP-eigenvector rank correlation is strong (Spearman rho = 0.934), but 39% of states diverge by five or more positions. Kentucky, for instance, ranks 28th in GDP but 14th in eigenvector centrality. Its commodity flows sustain production in wealthier states, yet this structural role is invisible in output-based accounting. Seven other physical-economy states show similar gains.

This is not merely an academic reranking. Seven of the eight overperformers are now attracting over $170 billion in AI data center investment (2024-2026). These states share characteristics that also attract compute infrastructure: cheap power, water access, grid capacity, and logistics corridors. A framework applied to 2017 trade data appears to identify structural assets the market has only recently begun to price.

We test boundary sensitivity by comparing the domestic network against a 52-node network including international flows. Eigenvector rankings are robust (rho = 0.982); betweenness centrality is not (rho = 0.816), raising open questions about weight semantics in flow networks.

ABSTRACT. Multidrug-resistant (MDR) ESKAPE pathogens exhibit remarkable robustness and adaptive capacity, yet the systems-level organizational principles underlying their interactomes remain poorly understood. Here, we present a comparative complex network analysis of protein–protein interaction networks (PPINs) across six major ESKAPE pathogens—Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter hormaechei—to identify conserved architectural features and functionally adaptive network components relevant to antimicrobial target prioritization.

High-confidence PPINs were constructed using STRING interaction datasets and analyzed using an integrated framework combining degree centrality, maximal clique centrality (MCC), K-core decomposition, and network cartography (z–P analysis). Two complementary prioritization strategies—KFC (K-core–Functional–Centrality) and FCK (Functional–Centrality–K-core)—were implemented to distinguish structurally conserved proteins from functionally adaptive network hubs.

Comparative analysis revealed a conserved dual-layer interactome architecture across all pathogens. The structural layer consisted predominantly of densely interconnected ribosomal proteins forming a highly conserved network backbone associated with cellular stability and survival. In contrast, the functional layer exhibited organism-specific enrichment in metabolic regulation, stress response, transport systems, signal transduction, virulence-associated pathways, and antimicrobial resistance mechanisms. Distinct adaptive signatures were observed across pathogens, including metabolism-dominated functional organization in Klebsiella pneumoniae and regulatory/stress-response enrichment in Acinetobacter baumannii and Staphylococcus aureus.

Importantly, structurally central proteins were found to possess limited therapeutic selectivity due to high conservation, whereas functionally central non-ribosomal hubs identified through FCK prioritization emerged as more promising candidate drug targets. Collectively, the findings support a systems-level model in which a conserved structural backbone supports pathogen-specific adaptive functional layers governing survival, resistance, and pathogenicity.

This study demonstrates the utility of comparative network science for uncovering emergent organizational principles in multidrug-resistant pathogens and provides a scalable framework for rational drug target prioritization in complex microbial systems.

ABSTRACT. Models of how things spread often assume that transmission mechanisms are fixed over time. However, social contagions--the spread of ideas, beliefs, innovations--can lose or gain in momentum as they spread: ideas can get reinforced, beliefs strengthened, products refined. We study the impacts of in branching processes using a model we call Self-Reinforcing Cascades (SRC).

Imagine a cascading product like a meme, conspiracy theory, rumor, or a piece of software spreading in a population of agents. At every transmission step in the cascade, the product has the chance to independently improve with probability $p$ or get worse with probability $1-p$. This process can stop for two reasons: either the quality of the product drops to zero, or the agents sharing it cannot find others to pass it on to.

The shape and statistics of SRC are surprising in two ways. One, they follow fat-tailed cascade size distribution in their subcritical regime (see Figure). Two, the tail of their depth distribution remains exponential even at the critical point. Thus, subcritical and critical SRC are short but very broad compared to classic cascade models. The unique shape of SRC could help identify signatures of varying cascade intensity or quality in empirical data. We could expect this mechanism to appear in cascades where content is able to change (e.g., mutations in epidemics or personalized social media posts) as opposed to cascade with fixed content (e.g., cascade of clones, or reposting in social media), and SRC could provide more realistic models in these cases.

We will discuss several important properties of self-reinforcing cascades and propose a few analytical approaches to better understand these processes: exact probability generating functions for cascade size, exact recursions over cascade depth, and the traveling wave technique for cascade intensity. Combining all three approaches, we are able to characterize the different scaling behaviors produced by the model. Altogether, this effort may provide a useful framework for researchers and practitioners seeking to understand cascading behavior in complex real-world systems.

ABSTRACT. Understanding emergence across levels—how micro-level interactions give rise to macro-level dynamics—remains a central challenge in complex systems science. While computational modeling has become a key approach for studying such systems, less is known about how different forms of interaction with models shape the development of micro–macro reasoning. In this work, we propose a design space for modeling-based environments defined by two dimensions: generativity, ranging from exploration of pre-built models to construction of models, and embodiment, ranging from disembodied interaction to embodied, participatory engagement. We argue that these dimensions function as systematic variations in how learners interact with the underlying mechanisms of a system, and therefore influence how emergence is understood. To examine this framework, we conducted a multi-condition study spanning four regions of this design space: normative instruction, computational model exploration, computational model construction, and embodied & computational model construction. Participants engaged with chemistry phenomena grounded in kinetic molecular theory, focusing on relationships between local interactions and aggregate system behavior. Learning was assessed in terms of micro-level reasoning, macro-level interpretation, and micro–macro integration, corresponding to key constructs in complex systems research. Results reveal a consistent ordering across conditions. Exploration supported recognition of macro-level patterns but showed limited gains in mechanistic reasoning. Construction led to stronger understanding of underlying interactions and improved ability to articulate causal mechanisms. The embodied construction condition produced the strongest performance overall, particularly in integrating micro- and macro-level explanations, suggesting that grounding local rules in embodied interaction supports understanding of emergent dynamics. These findings contribute to complex systems research by identifying generativity and embodiment as design dimensions that systematically shape reasoning about emergence. More broadly, they position modeling-based learning environments as intervention systems through which different forms of access to local interactions can be manipulated, offering a novel approach for studying how humans come to understand complex systems. This work suggests that understanding emergence can be systematically supported by structuring access to local interactions, providing a principled basis for designing environments that make complex systems more interpretable.

ABSTRACT. Progress in science and technology is punctuated by disruptive innovation and breakthroughs. To understand disruptive innovations and their drivers, the ability to operationalize and estimate “disruptiveness” is critical. Yet, this task remains difficult because scientific influence propagates through both direct and indirect citation paths, and discoveries are often fragmented across multiple papers. Here, we introduce an embedding-based metric of disruptiveness. When applied to large-scale publication data, the measure not only reliably identifies canonical breakthroughs, such as Nobel Prize–winning papers, but also finds simultaneous disruptions that eluded standard approaches. By enabling more robust identification of disruptive innovations and simultaneous discoveries, our method facilitates more accurate attribution of transformative contributions while providing insights into the mechanisms driving scientific breakthroughs.

ABSTRACT. Financial return distributions exhibit universal "stylized facts" such as fat tails and volatility clustering, suggesting that markets operate near a Self-Organized Critical (SOC) state where microscopic agent interactions cascade into large-scale price movements. Biondo et al. [1] reproduced this behavior through an order book model driven by trader imitation, but their framework assumes strictly independent time slice inherited from the typical sandpile model —an idealization that fails to capture the irregular timing of real decision-making. This study introduces two modifications into an OFC-based [2] order book: asynchronous updates, allowing each agent to act at independent intervals, and information decay, which gradually dissipates accumulated informational potential. Comparing four scenarios—(1) Asynchronicity + Information Decay, (2) Asynchronicity only, (3) Information Decay only, and (4) the original synchronous model [1]—we evaluate power-law behavior in avalanche size–frequency distributions and the CCFD of inter-event times. Only the fully integrated model (Scenario 1) consistently reproduces robust power-law behavior in both measures. Asynchronous updating drives the system toward a metastable critical state via "informational clogging," while information decay prevents saturation and sustains sensitivity to new information. Neither mechanism alone suffices; only their combination gives rise to the long-range correlations characteristic of volatility clustering.

ABSTRACT. How does an agent decide in a world whose "right" inference target is itself in flux? Classical Bayesian inference assumes a fixed likelihood P(d|h), so "rational" updating concentrates posterior on the data-supporting hypothesis; this fixity, however, fails in the open, non-stationary environments that biological cognition routinely navigates. Inverse Bayesian inference, introduced as a model of "natural-born intelligence" [1], answers this by letting the likelihood function itself be dynamically rewritten. Bayesian–Inverse-Bayesian (BIB) inference fuses conservative posterior updating with this generative reinvention of the inference targets: a process in which human-like flexibility self-organizes from self-rewriting likelihoods, manifesting empirically as Lévy-walk-like statistics with universal exponent ≈ 1.5.

We test this in reward-based BIB agents in symmetric N-hand rock-paper-scissors, a minimal substrate where CFR-family solvers collapse to uniform Nash play [2] and non-trivial dynamics originate from internal inference. The posterior trajectory P(h_t) is read through two dual observables: the argmax persistence T_argmax of the dominant hypothesis, and the laminar duration T_lam. At huge scale, BIB yields α_lam ≈ 1.34 and α_argmax ≈ 1.43, collapsing across a 2×2 design space and three opponent classes; the Bayes-only baseline, lacking the inverse step, is design-conditional. α_BIB is moreover invariant under hypothesis-count rescaling — criticality is internally self-organized, not parameter-tuned.

These results admit a clean interpretation as on-off intermittency in hypothesis space, in the sense of Platt–Spiegel–Tresser [3], whose canonical prediction for laminar-phase durations is P(T_lam) ~ T_lam^(-3/2). Bayes and inverse Bayes act at every step; laminar "off" phases of committed interpretation are punctuated by "on" bursts of reorganization. BIB thereby emerges as human-like flexibility self-organizing through the maintenance of robust criticality. The framework extends to multi-agent cyclic dominance and non-stationary environments.

References [1] Y.-P. Gunji, BioSystems 259, 105677 (2026). [2] N. Brown, T. Sandholm, Science 359, 418 (2018); 365, 885 (2019). [3] N. Platt, E. A. Spiegel, C. Tresser, Phys. Rev. Lett. 70, 279 (1993).

ABSTRACT. Nowadays, neural networks act as a synonym for artificial intelligence. Present neural network models, although remarkably powerful, are inefficient both in terms of data and energy. Several alternative forms of neural networks have been proposed to address some of these problems. Specifically, spiking neural networks are suitable for efficient hardware implementations. However, effective learning algorithms for spiking networks remain elusive, although it is suspected that effective plasticity mechanisms could alleviate the problem of data efficiency. Here, we present a new framework for spiking neural networks --- Spark --- built upon the idea of modular design, from simple components to entire models. The aim of this framework is to provide an efficient and streamlined pipeline for spiking neural networks (performance summary is shown subfigures a & b).

We showcase this framework by solving the sparse-reward cartpole problem without any auxiliary readout models and using a simple plasticity mechanism. Our solution consists of a simple A vs B model; activity on A is biased to inhibit activity on B and vice versa (subfigure c). Note that, Spark allows us to define self-recurrent complex architectures extremely easily. We trained 25 agents for a total of 1000 episodes each, followed by a testing of 250 episodes (subfigures d, e, f & g). The majority of our agents (16 out of 25) are capable of solving the problem within the first 50-80 episodes, while the remaining agents showed a slower training process; we allow agents to continue training for 1000 episodes to test the stability of our solution.

As shown with the cartpole example, we believe that a framework compatible with traditional ML pipelines may accelerate research in the area, specifically for continuous and unbatched learning, akin to the one animals exhibit.

ABSTRACT. Large language models (LLMs) are successful at capturing semantic content, as evidenced by their impressive conversational skills. However, it is not known whether their “semantic” representations capture the internal structure of concepts in humans, which is a key component of semantic organization in the human mind. One approach to analyzing conceptual structure is to represent an individual concept as a graphical network that captures the associations (edges) among its features (nodes). Here our goal was to use LLM embeddings to derive concept networks that are structurally similar to human-created networks. Fifteen concept networks were derived from LLMs and were compared to their human-derived counterparts in terms of structural network measures (e.g., modularity). For each concept (e.g., chocolate), Gemini was used to generate 50 sub-types (e.g., milk chocolate, chocolate sauce) and the embedding corresponding to each subtype was subsequently generated. Each concept network was created by assessing the similarities between feature vectors across sub-type embeddings for that concept. Five different similarity metrics were used for network construction (e.g., cosine similarity, Mahalanobis distance); we varied this similarity metric and swept across other parameters (e.g., network threshold, embedding dimensionality) to optimize structural correspondence between human and LLM-generated concept networks. Results show that concept networks constructed from Gemini embeddings successfully capture human semantic structure under several conditions. These findings suggest that LLM networks can capture the rich structure of human concepts. Further, these methods may be used to automate network creation for future research. The ability to automate the creation of human-like concept networks expands our knowledge of how large language models process text, and opens the door to further exploration of the flexibility of conceptual meaning.

ABSTRACT. The question of stability and longevity of governments has long attracted scholarly scrutiny. Svolik’s seminal work famously estimated that 42% of democracies are immune to authoritarian reversals, and his mixture model also implies that 77.2% of non-democratic spells end in democratization. However, methodological issues persist. First, there is a need for more refined, continuous measures of democracy, such as the Electoral Democracy Index, rather than dichotomous ones. Second, non-democratic regimes are prone to many destabilizing events (such as conflict, etc) and hence do not end through democratization alone. Third, hazard-based survival models, such as the one used by Svolik, tend to overestimate event probabilities -- a limitation that becomes crucial when the underlying decay dynamics are heterogeneous and driven by competing mechanisms. Hence, in this work, we introduce a competing risk survival framework to study the decay dynamics of non-democratic regimes in all countries between 1900-2021. By modeling democratization and three types of sovereignty loss— loss of international autonomy, recognition, and fiscal autonomy—as mutually exclusive exit events, we substantially revise prior democratization probabilities: given fiscal sovereignty, 52.2% of spells end in democratization, whereas 69% democratize given the state is recognized, and 49.8% democratize given international autonomy, compared to 77.2% estimated by Svolik’s approach. We further investigate how economic and institutional factors shape regime lifetimes and contribute to the observed variability. Our findings thus reveal that non-democratic regimes face a more fragile and complex landscape than previously understood. These novel insights show how democratization is shaped by competing instabilities, with broad implications for the study of regime dynamics.

ABSTRACT. Connectomics has provided insights into the organization of the human brain, demonstrating topological motifs like high connectivity among hub regions. However, detailed tracer-based and imaging studies in mammal brains show that the main feature determining connectivity between regions is not topology but rather spatial proximity. The exponential distance rule (EDR) strongly predicts connection strength between region pairs and applies also at the level of local cortical gray matter as well as in the projetome of Drosophila. EDR is thus a fundamental principle of brain organization and may reflect universal metabolic constraints. However, it does not predict all aspects of connectivity. Notably, it cannot explain asymmetries in bidirectional connection strength. To explain asymmetries, we suggest that brain organization is shaped not only by metabolic demands but also by the need to pass messages polysynaptically across a vast network whose topology is essentially fixed in adulthood, and do so while managing message-message interactions. We tested the ability of a variety of dynamical models and structural network measures to predict the largest absolute differences in directional connection weights on the mouse and marmoset connectomes. On the unweighted connectome, our copy-spread-annihilate (CSA) model generates message flux imbalances over edges that successfully predict empirical weight differences for those edges without any consideration of the brain’s embedding in space. We show that alternative models and measures either fail to explain weight asymmetries at all or predict less variance than CSA. The CSA model is a synchronous Markovian agent-based model with two key features: message broadcasting, which has been recognized as promoting message survival in brain networks, and is here modeled as duplication of injected messages to all neighbors; and message-message interactions (e.g., classical inhibition, XOR, and other gating mechanisms) which have been almost entirely ignored in the brain network communication literature, and are here modeled as mutually destructive collisions at nodes. Although the CSA model is abstract, and does not aim to capture biophysical interactions directly, it highlights the underappreciated problem of network-wide communication in the presence of message interactions. We conclude that we can go some way in explaining empirical brain organization purely in terms of network topology and message dynamics, as long as we take a sophisticated view of message routing demands.

ABSTRACT. Many network-based physical systems, ranging from aging spring networks [1] to electrical networks [2] and mechanical metamaterials [1], adapt their internal structure through purely local interactions to produce a wide range of functional behaviors, with no centralized control and no global signal indicating the desired output. In these systems, learning unfolds at the level of the edge parameters of a fixed underlying network. A variety of local rules and target functions have been studied extensively, but almost always on disordered networks that are derived from jammed packings. We argue that network structure strongly shapes the space of learned solutions. In the present talk, we examine this for the task of opening a phononic band gap on a disordered mass–spring network using a strictly local update rule. To our knowledge, this is the first physically realizable rule for adapting a global spectral property. All previous methods for opening band gaps have been global [3]. Whether the system succeeds depends heavily on the network structure (see Fig. 1). For example, spatial networks and Watts–Strogatz networks typically succeed, whereas Erdős–Rényi networks mostly fail. These results identify the underlying network as a fundamental determinant of what local-rule physical learning can achieve, motivating a graph-theoretical framework for how network structure shapes the space of learned solutions.

[1] N. Pashine, D. Hexner, A. J. Liu, S. R. Nagel, Sci. Adv. 5, eaax4215 (2019). [2] M. Stern, D. Hexner, J. W. Rocks, A. J. Liu, Phys. Rev. X 11, 021045 (2021). [3] H. Ronellenfitsch, N. Stoop, J. Yu, A. Forrow, J. Dunkel, Phys. Rev. Materials 3, 095201 (2019).

ABSTRACT. Mathematical modeling of complex systems is ubiquitous in modern science and engineering fields, and underpins advances in physics, biology, epidemiology, and weather prediction. As complex systems with a large number of variables and parameters, it is highly nontrivial to formulate and calibrate mathematical models to describe such systems; such models typically are formulated as nonlinear dynamical systems, which are challenging to understand both analytically and numerically. Recently, data-driven modeling has experienced a surge of research interest in the modeling community, due to its capability to effectively and efficiently learn rich mathematical structure from observations. However, a fundamental challenge remains: deciding whether the underlying process is inherently deterministic or stochastic. Indeed, randomness in measurements may arise intrinsically, may be due to unobserved processes, or may result from environmental fluctuations and/or experimental variability. Nevertheless, despite these mechanistically diverse mechanisms, the resulting data can appear statistically similar. Thus, a major open question is how one can construct and calibrate mathematical models when the source of randomness is unknown.

To address this challenge, we have developed a unified mathematical and computational framework that bridges the learning of deterministic and stochastic dynamical systems. Traditionally, when calibrating models, one makes a priori assumptions on the model structure as either purely deterministic or stochastic. However, many systems possess hybrid singular diffusion structures, where randomness acts directly on certain components only; deterministic systems arise as a special class where the diffusion matrix is completely degenerate. The objective of this work is to create unified learning methodologies that are model agnostic, and are thus able to predict, from the observational data alone, the equations and the degree of stochasticity present in the underlying physical process. More specifically, we have developed a noise-informed variational learning framework (based on quadratic variation for the noise and Girsanov's theorem for the drift), and have shown analytically that the resulting estimator is consistent and asymptotically normal under relatively weak assumptions. We also develop a Q-weighted framework to learn the drift, and show that the resulting minimizer is invariant with respect to (symmetric positive definite) changes in the covariance-based weighting; however, a canonical choice based on the diffusion covariance structure is optimal with respect to asymptotic efficiency. The developed methodology is also validated on a broad class of models from physics, biology, and finance, including a high-dimensional stochastic Cucker-Smale flocking model and an mRNA bursting chemical Langevin equation, highlighting the fact that the method remains effective across state-dependent, correlated, singular, and high-dimensional systems. Output for learning the reaction rates in a model of mRNA bursting is provided below; trajectory data distributions utilized are also plotted (blue histogram).

ABSTRACT. In their classic book, Hertwig and Engel [1] define deliberate ignorance (DI) as the conscious choice not to seek or use available information. Their volume develops a taxonomy of DI functions and concludes that DI is a fundamentally social phenomenon, but quantitative work on the population-level conditions under which information is accepted or rejected remains limited. This paper presents a cellular automaton model and a parameter sweep aimed at identifying the dominant drivers of informational uptake and their interactions. We build a stochastic cellular automaton on a 150 x 150 toroidal grid populated by 22,500 agents in three states: uninformed, informed (absorbing), and stubborn (a fixed 1% boundary case). At each step an uninformed agent is exposed to information through its Moore neighbors with probability p_social * n_i, where n_i is the local share of informed neighbors. If exposed, the agent accepts with probability clip(beta_0 - beta_risk * event_risk + beta_social * n_i, 0, 1). We interpret beta_0 as a dispositional parameter (baseline acceptance), beta_risk * event_risk as the risk-suppression mechanism associated with emotion-regulation avoidance, and p_social together with beta_social as social-environment factors. We sweep all five parameters at four levels each (1,024 combinations), with 8 replicates per combination, for a total of 8,192 simulation runs. Combination-level means are summarized using linear models with pairwise interactions, lasso regression, and decision trees. The final mean informed share spans nearly the full unit interval (0.10 to 0.99) across the parameter space. Three results follow (Fig. 1). First, beta_0 is the dominant driver: increasing it from -0.30 to +0.15 raises the mean informed share from 0.10 to 0.79. Second, event difficulty produces a roughly fifty-percent reduction in informed share between low- and high-difficulty conditions (0.51 to 0.26). Third, social factors alone are insufficient to overcome strong refusal dispositions: marginal means over the p_social x beta_social plane span only 0.20 to 0.45, compared with 0.10 to 0.95 over beta_0 x event_risk. Statistical summaries account for 72% - 80% of cross-combination variance, and every leading pairwise interaction term involves beta_0 (decision-tree importance: beta_0: 0.55, event_risk: 0.17, beta_risk: 0.14, beta_social: 0.09, p_social: 0.04). Population-level deliberate ignorance in this model arises from the joint configuration of disposition, event difficulty, and social environment, providing a quantitative complement to the qualitative account in [1]. [1] R. Hertwig and C. Engel, eds., Deliberate Ignorance: Choosing Not to Know, Strüngmann Forum Reports, vol. 29, MIT Press, 2020.

ABSTRACT. A central question in complex systems science is how non-trivial structure emerges from simple local interactions, and how system topology—such as network structure—shapes the outcome. Reaction-diffusion systems, spatial games, and ecological models have repeatedly shown that spatial patterning can enable the exploration of state spaces inaccessible to well-mixed dynamics. This raises the question of whether spatial patterning could also impact combinatorially open chemistries, a critical issue for explaining emergent complexity. Can spatial structure alone—in the absence of any explicit selection—guide a chemical network toward objects of higher structural complexity?

We investigate this question using a model of computational chemistry based on a network of chemical reactors within which linear polymer chains are represented as integers that can undergo synthesis and decomposition. Our model retains core properties of molecular chemistry—composition and open-endedness—and includes key transport parameters of inflow, diffusion, and outflow. We use the assembly index to assess the effects of varying the parameters and topology of the system on the complexity of individual objects [1,2].

Our analysis reveals distinct regimes, controlled by diffusion and inflow, that impact the formation of high-assembly objects: a local homogeneous regime where most of the material remains localized near the source, a heterogeneous regime where dispersion suppresses second-order constructive reactions, and a global well-mixed regime at high diffusion, which caps the achievable complexity but maximizes the exploration of system topology. As network exploration increases, we show that topology shapes high-assembly objects: ordered structures, such as lattices, promote the formation of more complex species than degree-preserving randomized controls (Fig. 1)—an effect that first-passage-time analysis shows exceeds what residence time can account for. Topology, therefore, acts as a tunable parameter for emergent complexity, and provide a genuine form of implicit selection for complex objects.

Our results highlight the critical role of both transport processes and network topology in shaping complexity. They demonstrate how complex objects might emerge within chemical systems in the absence of explicit selection, with implications for prebiotic chemistry and the emergence of biological complexity.

[1] Marshall et al., Nat. Comm. 12, 3033 (2021). [2] Marshall et al., Entropy 24, 884 (2022).

ABSTRACT. Empirical research on mindfulness has relied on reductionist decomposition, by isolating and analyzing discrete components and mechanisms of mindfulness mechanisms. While valuable, this approach is poorly suited to capturing how change and development emerge from continuous, nonlinear, and recursive interactions unfolding across multiple timescales. Through a dynamic systems (DS) approach, we aim to empirically investigate the attractor/repellor state space of mindfulness meditation. Specifically, we examine the objects of mindful awareness as well as the temporal dynamics of mindful awareness and mind wandering during meditation using a novel behavioral assay (the Mindful Awareness Task or MAT) among 143 primarily meditation-naïve participants. First, we aim to characterize (1) patterns of the phenomenological modality (sensory vs. mental) and the hedonic tone (pleasant vs unpleasant) of the objects of attention during mindful awareness; and, second, to investigate (2) the inter-relations between these patterns and the temporal dynamical features (e.g. sustained mindful awareness during meditation, latency to return to mindful awareness following mind wandering). Our first study found evidence for two patterns of bias in mindful awareness during meditation – a sensory bias and a hedonic negativity bias (see Figure 1). First, among all objects of mindful awareness with an unpleasant hedonic tone, we found a significant bias toward sensory experiences (78.2%) relative to mental (21.8%). Conversely, among all objects of mindful awareness that were sensory, we found a significant bias toward unpleasant (95.5%) relative to pleasant hedonic tone (4.5%). Second, we found that greater levels of the observed sensory bias were significantly associated with greater overall time in sustained mindful awareness during meditation (b = 176.66, SE = 79.73, p = .03) and shorter latencies to return to mindful awareness after periods of mind wandering (b = -5.83, SE = 1.77, p = .001), and that greater levels of the hedonic negativity bias were also associated with shorter latencies to return to mindful awareness (b = -4.65, SE = 2.23, p = .04). We theorize that these dynamical features of mindful awareness may reflect the presence of attractors and repellors in the state-space landscape of meditative experience. Thus, to more directly quantify the state-space landscape (e.g., attractors/repellors, dispersion, and entropy) of the temporal dynamics associated with these sensory and hedonic negativity biases of mindful awareness, we will apply state-space grid analysis to the MAT data. Finally, we will discuss implications of the findings for understanding how mindful awareness can be cultivated over micro- and macro- timescales.

ABSTRACT. Real complex systems exhibit rich collective behavior, yet identifying which components of an interaction network drive such dynamics remains a central challenge. Here, we show that complexity itself can resolve this problem. In large random and empirical networks, structural disorder and heterogeneity induce spectral localization, causing Laplacian modes to concentrate on small subsets of nodes. This converts global modes into identifiable dynamical units tied to specific structural components. Exploiting this principle, we develop a node-resolved stability framework that predicts instability onsets, identifies the nodes responsible for collective transitions, and restores interpretability in systems where classical modal theories fail. In heterogeneous reaction networks, the same mechanism enables collective states beyond those usually associated with homogeneous assumptions. More broadly, our results show that complexity can be revealed, rather than obscure, the microscopic drivers of macroscopic dynamics.

ABSTRACT. The analogy between rogue waves in hydrodynamics and optics was first spurred on by the recognition that the wave heights or intensities in both systems followed a long-tailed distribution [1]. Here, we extend this analogy to financial volatility indices and show that by identifying peaks of the envelope function of a volatility index, extreme waves exceeding the typical significant wave height threshold used in ocean physics can be identified. These peaks can be seen for the CBOE VIX in the top panel of Figure 1 a). The values of the volatility at the peak locations will follow a long-tailed distribution as seen in the middle panel of Figure 1 a), however the prominences of these waves show a power-law decay between occurrence and height, see the bottom panel of Figure 1 a). We use the eigenvalues and eigenvectors of the linearised nonlinear Schrödinger equation to examine the behaviour of the system before an extreme event. To do this, we extract the envelope wave of the system only up to the last available datapoint and then calculate the numerical gradient of the minimum eigenvalue. The upper panel of figure Figure 1 b) shows the maximum eigenvalue gradient found in the ten days up to and including the point at which it is displayed, and the lower panel shows our rogue wave warning indicator, which counts the number of signals exceeding an arbitrary threshold in those ten days, and combines it with the maximum eigenvalue gradient. Once enough data is present to resolve the envelope wave, the signals arrive up to 20 days before a peak (depending on the timescale of the event), typically only missing events which happen over timescales too short to capture in advance, such as the peak in 2018. This analogy may therefore provide a new source of early warning indicators for extreme financial events.

ABSTRACT. Contentious political processes display the hallmarks of complex dynamical systems: heterogeneous causal mechanisms acting across multiple ontological layers (actor, network, structure, field), temporal reweighting of those mechanisms, and emergent macro-outcomes irreducible to any single component. Yet social science remains fragmented into rival schools, each focusing on its favored causal mechanism. Social science needs a common metric for weighting the relative and conjoint causal contributions of many mechanisms in a specific historical case. This paper presents Integrative Structurational Analysis (ISA) and tests it on a 27-year environmental conflict in Oita, Japan (1955–1982). By analogy with physical systems — where local events emerge from a small set of fundamental forces under varying boundary conditions — ISA reconstructs the social field along three orthogonal causal dimensions: malleability (3 levels of structure–agency determination), tangibility (12 media of power, from coercive violence to moral-ethical and biophysical), and mechanism (20 dyadic interaction types including domination, brokerage, diffusion, scale shift, and structural cascade). These axes define a 720-cell power cube in which each cell represents an elementary mode of causal influence between units (actors, actants, or fields). Empirically, the macro-process is decomposed into hundreds of micro-interaction dyads, each scored on the power relational dimensions and weighted by its impact on subsequent system states. Aggregation across nine successive phases of just the tangibility data (in 6 categories) reveals shifting causal configurations — exposing how factors emerge, magnify, and diminish to produce non-obvious outcomes (see Figure 1). (Figure 1 about here-see uploaded pdf copy) Figure 1. Causal configuration timescape of the Oita case. Nine three-year phases (1955–1982); y-axis is total causal weight summed over outcome-influencing micro-dyads. Annotations mark the dominant mechanism in turning-point phases. Total system activity peaks in phase 6 (symbolic interaction and resistance) and relaxes by phase 9, where capital withdrawal under global recession (agentic-economic) produced an unexpected outcome inverting the trajectory set in phases 1–3. Six-medium projection adapted from Broadbent (2024); the present paper extends the analysis to the full 12-medium tangibility axis.

ABSTRACT. We investigate continuous-time random walks (CTRWs) on undirected networks and show that several dynamical properties exhibit universal behavior independent of the specific form of the waiting-time distribution. We demonstrate analytically and numerically that the steady-state occupation probability of a node depends solely on its degree and is independent of the waiting-time distribution. In contrast, the relaxation toward the steady state is controlled by whether the mean waiting time $\langle \tau \rangle$ is finite. For heavy-tailed waiting times $\psi(\tau) \sim \tau^{-1-\alpha}$ with $0 < \alpha < 1$, the relaxation inherits $\alpha$-dependent anomalous behavior, while for $\alpha > 1$ all CTRWs collapse onto a universal relaxation curve determined only by network structure. We show that recurrence statistics reflect the long-time tail of $\psi(\tau)$, and that the mean first-arrival time exists only when $\langle \tau \rangle$ is finite. Finally, we analyze the exploration dynamics and find that the mean number of distinct nodes visited grows algebraically, $\langle S(t) \rangle \sim t^\alpha$ for $\alpha < 1$, whereas for $\alpha > 1$ it exhibits the universal scaling $\langle S(t) \rangle \sim N f(t / (N \langle \tau \rangle))$, analogous to discrete-time random walks. These results reveal that the existence of a finite mean waiting time partitions all CTRWs on networks into two universal dynamical classes, irrespective of waiting-time details.

ABSTRACT. Modern economic activities operate as complex international systems in which institutional interventions inϐluence social outcomes indirectly through global supply chains. Regulations such as the Kimberley Process Certiϐication Scheme (KPCS) and Section 1502 of the U.S. Dodd–Frank Act (DFA1502), which target conϐlict minerals through supply chain due diligence and transparency requirements. Despite their growing importance, empirical evaluations of how these interventions propagate through multi-stage causal processes remain limited, particularly in internationally comparable and long-term contexts. This study examines the causal chain linking international regulations, economic structures, and downstream social outcomes. In this study, KPCS and DFA1502 are conceptualized as exogenous regulatory shocks that alter upstream trading rules and compliance requirements. These changes affect intermediate economic variables including trade patterns, transaction structures, and revenue allocation which may subsequently inϐluence downstream conϐlict dynamics and socio-economic instability. To evaluate these processes, the study develops a counterfactual causal inference framework using cross-national panel data. Methodologically, it employs the generalized synthetic control method to estimate counterfactual trajectories while accounting for latent global shocks and heterogeneous country-speciϐic responses. The framework further incorporates mediation analysis to examine how institutional interventions transmit their effects through intermediate stages of global supply chains. In addition, temporal dynamics and lagged responses are modeled to assess how institutional effects gradually emerge and diffuse over time. By conceptualizing international institutions as partial interventions within complex socio-economic systems, the study proposes a structured empirical framework for analyzing how regulatory interventions propagate through global supply chains. The approach contributes to methodological discussions at the intersection of causal inference, complex systems modeling, and international political economy.

ABSTRACT. Perceptions of how common an attribute or behavior is play a central role in shaping individual decisions and, in turn, collective outcomes. What we perceive depends on what we see, which itself is shaped by the social network we live in. Existing literature quantifies prevalence as the fraction of one-hop neighborhoods that share a given attribute, but in reality, we are also exposed to the behavior of friends-of-friends and beyond.

We introduce a multiscale framework for perceived prevalence in which a diffusion process propagates exposure across the network. Using a continuous-time random walk, the perceived prevalence of node~$i$ at scale~$\tau$, denoted as \(\phi_i(\alpha, \tau)\) for a binary attribute ~$\alpha$ such as smoking, is the fraction of total probability of random walkers originating from attribute-holding nodes that reach~$i$ by time~$\tau$. In the limits $\tau \to 0$ and $\tau \to \infty$, this recovers the local and global prevalence, respectively. We apply the framework to a synthetic and a real-world network, showing that it uncovers mesoscale structures that are invisible to local or global measures. Notably, we use the intermediate-scale perceived prevalence to predict smoking uptake in the subsequent wave and show that it outperforms local prevalence as a predictor. This is particularly valuable for estimating the impact of structure beyond one-hop neighborhoods on behavior adoption.

ABSTRACT. This study employs recurrence quantification analysis techniques to investigate metastable and hysteretic non-linear dynamics within Ecological Momentary Assessments (EMA) data, which capture real-time psychological states in subjects’ natural environments. To identify metastable states, we extended a procedure proposed in [1], based on Recurrence Plots with the spatial radius ε estimated via Markovian optimization. We correlated these results by constructing a dynamic graph, where edges represent correlations in item changes between consecutive measurements. Validating our technique, we observed a high correlation between node degree and the number of meta-states, which significantly decreases during the subject’s transition into depression [2]. For the identification of the hysteretic regime, given the strongly non-linear nature of the phe- nomenon, a novel algorithmic pipeline was designed. The process begins by identifying a direc- tional relationship between the items and subsequently calculating the time lag between them, thus addressing the problem of delayed effects intrinsic to psychological data. Following the temporal re-alignment of the two series, we generated the bifurcation diagram by dividing the parameter values into increasing and decreasing sub-series, color-coding the respective variable values. This approach allows for the study of the variable’s dynamic evolution as a function of its history. The entire procedure was validated by constructing an artificial dataset simulating the hysteretic system ˙ x = r + x− x3, applying a time shift between the parameter r and the variable x. As shown in figure 1, our pipeline proves capable of accurately identifying the inserted lag, correctly distinguishing between variable and parameter, and reconstructing the exact bifurcation diagram.

[1] P. Beim, Graben and A. Hutt. Detecting metastable states of dynamical systems by recurrence- based symbolic dynamics. Physical Review Letters, 110:154101, 2013. [2] I. Caporossi, S. V. George, and C. Mocenni. Metastability from recurrence analysis in depres- sion. Manuscript submitted for publication, 2026

ABSTRACT. Understanding human mobility is essential for capturing urban activity patterns and designing mobility systems. POI (Point-Of-Interest) networks provide an effective means to quantitatively analyze such patterns and have been shown to improve the accuracy of economic forecasting [1]. However, most studies focus on large metropolitan areas, with limited attention to regional cities or temporal behavioral changes.

In this study, we constructed and analyzed a POI network using anonymized GPS data collected in 2024 for a specific region in Japan. POI data were obtained from the publicly available dataset provided by Foursquare. We then compared mobility changes between holidays and weekdays. We also categorized visited POIs into lower-order and higher-order needs [2], and transitions between categories were represented as a network.

The results suggest observable differences in the mobility patterns of individuals. On weekdays, mobility is more concentrated in lower-order needs, whereas on holidays, transitions toward higher-order needs become more prevalent (Fig. 1). This suggests that the relaxation of temporal constraints shifts human behavior from necessity-driven activities toward those driven by preference and psychological fulfillment.

Future work will evaluate the impact of mobility optimization in autonomous driving on well-being from a network perspective, providing insights into the social implementation of mobility systems that support well-being.

ABSTRACT. Traffic assignment, i.e. route and travel-time choice and their impact, constitutes an essential element in transportation planning and management. Solving the network traffic assignment makes it possible to predict the load of traffic in the network, and thus to assess the consequences of planning (long term) and management (short-term) measures. The traditional approach to traffic assignment is based on the utility concept: travelers are assumed to choose among all available alternatives (routes, departure times) the one which is most advantageous, i.e. the alternative with maximum utility. The reference [1] reviews many recent progresses in traffic assignment. In this presentation we propose a different approach, based on the MFG (mean field game method) [2,3]. The idea of our approach is that travelers adapt their choice progressively, on a day-to-day basis, in order to minimize their travel cost. During this dynamic process, travelers also want to minimize their rate of change of choice, since they deduce their choices from experience and prediction. The travel costs perceived by travelers result from traffic conditions, which in turn result from the choices of all travelers. The interaction between the optimization of the individual choice processes and the traffic state can be described as a MFG. References such as [3,4,5] address traffic assignment and MFG, under the route optimization angle with no reference to day-to-day choice. Our approach encompasses route and departure time choice on a day-to-day basis, describes the travelers' collective behavior and accounts for the progressive equilibrium formation [6]. Figure 1 shows, as function of time, the final departure time distribution (green, re-scaled) and the convergent evolution of congestion queues in the network (turquoise, one curve per day).

ABSTRACT. Foreigners are often associated with an increasing level of violence and insecurity. Yet, the links between crime and migration are unclear and difficult to observe because of the many overlapping and interconnected components [1]. For a variety of reasons, such as fear of deportation or a lack of a solid criminal network, foreigners are usually less likely to commit crimes than similar natives. Yet, even if foreigners might be less likely to participate in crime, they tend to be disproportionately present in jail [2]. In Switzerland, for example, foreigners account for 71% of the prison population despite comprising only 23% of the total population, while in Austria, of the 10,000 people in prison, about 53% are foreign-born, despite their relative population share of only 22%.

This study examines why foreigners may be disproportionately represented in prison despite evidence that they may be no more likely, and often less likely, to commit crimes than comparable nationals. We develop an agent-based model for the criminal justice cycle, considering crime participation, conviction, sentencing, incarceration, sentence reduction, release, and reoffending. The model is then calibrated to the Austrian criminal parameters (so, the number of offences, convictions, foreigners and so on). By varying parameters for nationals and foreigners, we show that modest differences in conviction rates, incarceration probabilities, and sentence lengths can substantially increase the share of foreigners in prison, even when criminal behavior is equal across groups. The results suggest that prison statistics may reflect unequal enforcement and sentencing processes rather than differences in offending alone.

[1] Brian Bell and Stephen Machin. The impact of migration on crime and victimisation. Centre for Economic Performance, London School of Economics, 2011 [2] Robert J Sampson. Rethinking crime and immigration. Contexts, 7(1):28–33, 2008.

ABSTRACT. AI professors (perfectly rational, inexhaustible) increasingly replace human teachers, yet the strategic interaction with boundedly rational, effort-averse human students has received limited formal game-theoretic attention. We fill this gap with a two-stage signaling game of incomplete information about student ability [2]: the AI chooses a teaching signal s (difficulty/feedback intensity) with cost c s^2; the student observes s then chooses effort e; private ability θ ∈ {L,H} affects effort cost k_θ e^2. Student best response is e*(s,θ) = β s / (2 k_θ); if the AI knew θ, optimal signal would be s*(θ) = α β θ / (4 c k_θ). Under incomplete information we derive two perfect Bayesian equilibria: (i) a pooling equilibrium (low AI cost c) where both types receive the same signal, producing a "learned helplessness"-like outcome that crowds out high-ability students; (ii) a separating equilibrium (intermediate c) where signals reveal ability, improving welfare. A key result is non-monotonic student welfare as a function of c: very low c (pooling) and very high c (AI stops teaching) reduce welfare, while intermediate c yields separation and maximal learning. Agent-based simulations (Figure 1) confirm the theoretical thresholds (c < 0.2 pooling, 0.2 < c < 0.8 separation, c > 0.8 collapse). Over-personalization (low c) can be harmful; AI may benefit from "strategic opacity" – calibrated difficulty to screen ability – a design principle suggested by signaling theory and economic analysis of AI [1,2].

[Insert Figure 1 here]

Figure 1: Student expected utility vs. AI cost parameter c. Low c: pooling → low effort. Medium c: separating → high effort. High c: collapse. Parameters: α=β=1, p=0.5, θ_H=1, θ_L=0.5, k_H=0.5, k_L=1.5.