ABSTRACT. Autonomous systems increasingly require moral judgment capabilities, yet whether these capabilities scale predictably with model size remains underexplored. We systematically evaluate 75 large language model (LLM) configurations spanning 0.27B to 1000B parameters using the Moral Machine framework [1]. For each model, we compute Average Marginal Component Effect (AMCE) vectors across nine moral factors and measure alignment with aggregate human preferences as the Euclidean distance D from the human AMCE vector. We observe a consistent power-law relationship D ∝ S−0.10 ± 0.01, where S is model size in parameters (Fig. 1) [2]. This power-law formulation outperforms linear, logarithmic, and exponential alternatives. Beyond improved mean alignment, variance in alignment decreases with model size, indicating more reliable and consistent moral judgment at scale. To assess robustness, we fit linear mixed-effects models with model family as random effects, controlling for release date and reasoning capability. Release date did not improve model fit, whereas extended reasoning (chain-of-thought and thinking-mode models) significantly improved alignment. Critically, a significant size × reasoning interaction shows that extended reasoning benefits are most pronounced in smaller models, while very large models already capture much of this capability through scale alone. After controlling for all confounds, model size remained a significant predictor. These findings extend scaling law research from tasks with objective ground truth to value-based judgments, demonstrating that moral alignment follows predictable power-law patterns across diverse architectures and training approaches. The modest exponent (α ≈ 0.10) suggests that moral judgment is a slowly emerging capability requiring substantial computational scale, with complementary benefits from extended reasoning architectures. These results provide quantitative foundations for risk-based deployment of LLMs in ethically sensitive applications.

ABSTRACT. Electric-vehicle (EV) adoption turns urban charging into a coupled mobility–energy system: individually reasonable charging decisions can collectively create localized overloads across distribution feeders. Using the San Francisco Bay Area as a case study, our central question is: what is the maximum system-level benefit of coordinated EV charging for mitigating feeder overloading risk?

We develop a mobility-aware coordination framework that models each EV as an agent embedded in a week-long activity trajectory. Unlike approaches that impose session-level energy targets, our formulation only requires each vehicle to maintain enough state of charge for future trips. This creates a massive optimization problem with temporal coupling from battery dynamics and spatial coupling from vehicle movement across feeders. We solve it with a custom ADMM-based distributed algorithm for parallel clusters, supported by theoretical analysis and empirical benchmarks.

Using realistic mobility trajectories and feeder-level hosting-capacity data, we optimize one-week charging schedules for roughly 2 million EVs across 1300+ feeders under a 30% adoption scenario. Compared with rule-based or individually optimized charging, full coordination nearly eliminates feeder upgrade needs, saving billions in potential costs. Further constrained and limited-coordination experiments identify myopic session demand targets as a key barrier to realizing the collective flexibility of EV fleets.

A preprint that details the methodology is available at: https://arxiv.org/abs/2604.11999

ABSTRACT. How groups reach collective decisions is central to understanding human and animal behaviour and to designing decentralised artificial systems such as robot swarms or networked AI agents. Most models of opinion dynamics assume that individuals update their opinions instantaneously upon social interaction. However, opinion change often involves a finite deliberation time during which individuals evaluate new information before adopting and transmitting it. Accounting for such delays can fundamentally alter collective dynamics. Here, we study collective decision-making in time-varying networks where individuals require a finite time to deliberate before switching opinions. We show that the temporal scale of network rewiring critically determines whether new opinions can spread through the group. In particular, rapidly changing social networks hinder the global adoption of new opinions, especially when deliberation times are long, despite new opinions spreading faster than the old ones at the local level. Conversely, sparse and slowly evolving networks reliably support the spread of superior options, largely independently of individual deliberation time. Our results reveal a fundamental trade-off between network volatility and cognitive processing time, providing new insights into how social and artificial systems can balance flexibility and stability to enable effective collective decision-making.

ABSTRACT. U.S. universities have rapidly produced policies governing generative AI in academic integrity, admissions, teaching, and institutional technology use, yet little is known about how task-force composition shapes what these policies emphasize. We assembled 7,609 AI governance documents from over 1,200 U.S. higher-education institutions, including all R1 universities and a stratified IPEDS sample, alongside rosters of institutional AI task forces coded by functional role. Topic modeling (LDA, K=15) reveals that defensive topics—academic dishonesty, AI-generated content and citation, and data security—dominate at 69.1% of policy documents, while integrative topics such as GenAI literacy, teaching strategies, and workforce preparation account for only 22.8%. Bayesian beta regressions linking task-force composition to topic prevalence show that Senior Leadership presence is associated with greater emphasis on plagiarism and AI-generated content and lower emphasis on GenAI literacy and data security, while Humanities/Social Science representation predicts greater focus on citation and authorship. Education/Teaching faculty and students are systematically underrepresented on task forces. Drawing on upper-echelons theory, we argue that university AI policy is not neutral technocracy: who writes the rules shapes what the rules emphasize, with consequences for how the next generation engages with AI.

ABSTRACT. This work proposes a new hybrid opinion model, CoDiNG (Continuous-Discrete Naming Game) [1], which combines discrete and continuous approaches to opinion modeling. The continuous part of the model has the potential to capture the complex cognitive processes of the human brain involved in opinion formation, while the discrete part represents the verbalization of the opinion during interactions with other people. This mechanism brings the model closer to the real dynamics of opinion formation and is grounded in cognitive science. Our CoDiNG model is based on the Naming Game but extends it by incorporating a mechanism from the CogSNet model [2]. Similar to the classical Naming Game, our model distinguishes three states (opinions): agree, disagree, and not sure. The CogSNet mechanism introduces an additional hidden continuous layer that represents the strength of belief in a specific opinion. Each node possesses its own opinion and corresponding weights for each of the three states (belief strength). Nodes interact with each other, modifying the weights of other nodes, leading to opinion changes when the difference between weights exceeds a predefined γ threshold. The mechanism for updating opinion weights, including their reinforcement through interactions and weakening over time in the absence of interactions, is consistent with cognitive science insights into memory traces in the human mind. The CoDiNG model was tested on data from the NetSense longitudinal study [3], utilizing the communication history and detailed surveys completed by participants as a unique real-world ground truth. Among hundreds of survey questions, we selected six that exhibited the greatest variability in opinions, covering topics such as euthanasia, marijuana, and various social benefits. Opinions from the first survey completed by participants were used to initialize the model. We then conducted simulations in which nodes communicated based on the participants’ communication history, leading to changes in their opinions. Finally, we compared the CoDiNG model’s output with the opinions of participants from the most recent survey they completed. Our experiments show that, in many cases, the CoDiNG model outperforms the classical Naming Game baseline by 20-30%. We observed that the continuous-discrete mechanism works better for general societal issues compared to highly individual matters, such as euthanasia. We also found that the value of the γ parameter is crucial, and for each survey question, the optimal value lies between 0.2 and 0.3. The proposed model and the obtained results offer extensive interpretive possibilities for the social sciences.

ABSTRACT. Opinion dynamics in social systems emerge from local interactions among heterogeneous agents and evolving network structures. Understanding how these interactions shape collective behavior remains a key problem in complex systems. In this work, we propose a hybrid cellular automata model that integrates heterogeneous agent roles, including normal users, stubborn individuals, influencers and mass media. The model includes self-confidence, homophily, and adaptive rewiring, enabling the co-evolution of opinions and network topology. Results show that self-confidence and homophily shape opinion clusters and network structure. Stubborn agents promote the persistence of diverse opinions, while media and influencers significantly impact global opinion. These findings provide insight into how local interaction rules drive opinion dynamics in adaptive social systems.

ABSTRACT. Collective opinion formation under uncertainty is a central problem in complex systems and computational social science. Individuals often belong to multiple social groups such as families, workplaces, and local communities, and they are frequently influenced by these groups at the collective level. Such influences are more naturally represented as higher‑order interactions on hypergraphs rather than simple pairwise links [1]. However, individuals are not blindly influenced by every group to which they belong. Instead, when individuals belong to multiple groups that exhibit different opinion trends, it is natural to assume that they compare the information obtained from these groups and evaluate which groups are more credible based on the observed opinion distributions, and then adjust their opinions accordingly.

Motivated by this perspective, we study an opinion dynamics model on hypergraphs where agents evaluate the credibility of the groups to which they belong. Credibility depends on two indicators: the internal consistency of opinions within the group and the deviation of the group’s opinion distribution from the surrounding opinion environment. Agents update their opinions by integrating information from multiple groups while adjusting the influence from each group according to its credibility. The parameter η is a learning rate controlling credibility weight updates; simulations on 3 uniform complete hypergraphs reveal a sharp qualitative change in collective behavior as η increases. For small η, the system converges to consensus, whereas above a critical learning rate scale η_crit opinions form persistent clusters, which can be interpreted as an effective disconnection of hyperedges induced by credibility learning. Linear stability analysis around the consensus state provides a theoretical estimate of this threshold consistent with simulations. We also discuss directions for understanding collective opinion formation in real-world social systems through further theoretical analysis and by examining opinion dynamics on various hypergraph structures.

ABSTRACT. Collective decisions in social and engineered systems are shaped not only by interactions among agents, but also by intrinsic constraints within individuals such as adherence to belief systems in social networks and hardware limitations in autonomous robots. To understand effects of such heterogeneity on group decision outcomes, we study a model of dynamic decision-making on multiple alternatives based in the recently proposed nonlinear opinion dynamics (NOD) framework. Collective decisions in NOD networks emerge in a bifurcation modulated by a tunable attention parameter u as multistable equilibrium opinion configurations. We incorporate agent limitations into NOD as orthogonal projection constraints, which restrict each agent’s opinions on multiple options to a feasible subspace. In the presented work, we show that projection constraints induce an invariant space on which the networked decision dynamics admit a reduced description. Analysis of the reduced system yields insights that heterogeneity in a group can make an unweighted communication network behave effectively like a network with weighted edges. In particular, we show that differences in hidden constraints on agents' beliefs redistribute influence and alter group sensitivity to inputs, leading the group to reach opposite collective decisions despite identical communication network structure and external cues about option quality. Hence, observed communication network structure alone may be insufficient to predict collective outcomes in heterogeneous groups.

ABSTRACT. We study pairwise bounded confidence opinion dynamics on a scale-free network that is steadily growing with the arrival of new agents. The probability that an arriving agent connects with a target agent of the network increases exponentially with target agent degree and opinion proximity between the two agents. Simulations show that, for a wide range of parameter values, cascades of opinion appear regularly (see Figure 1). Minor clusters form on the periphery of the opinion space and remain stable for a while, before merging more or less progressively with a major cluster in the network. These opinion cascades are absent when the network is totally connected [1]. We analyze these opinion cascades. They are triggered by the arrival of agents acting as "bridges" between the minor and major clusters, initially further apart than the confidence bound. While the minor cluster is further than the confidence bound from the major cluster, the move of the minor cluster towards the major cluster is slow. However, when the distance between the clusters crosses the confidence bound, the merging process becomes much faster. When the opinion homophily of arriving agents increases, the cascades become slower overall. We also study and explain this effect. These dynamics could represent stylized features of growing opinion groups or echo-chambers in online social media.

ABSTRACT. Models of opinion dynamics propose how individual interactions shape collective belief, but the field has long suffered from a dearth of empirical evidence linking models to observed behavior [1]. Three problems explain the gap. Empirical work that fits generative models typically restricts itself to a single parametric family on a single dataset, precluding both direct comparison between mechanisms and assessment of whether findings generalize across settings. And traditional agent-based calibration matches simulated statistics to stylized facts, which is not identifiable: very different microscopic rules can produce similar macroscopic behavior [1]. We address all three with a likelihood-based framework for inferring pairwise interaction kernels on a known network. The model is generative: at each step an agent picks a neighbor uniformly at random and shifts its opinion through a kernel k(xi,xj)plus Gaussian noise. The selected neighbor is unobserved, so we marginalize it with expectation-maximization. The framework supports four classical parametric kernels (DeGroot, Hegselmann-Krause, sigmoidal bounded confidence [2], and a smooth attractive-repulsive kernel; Fig. 1a) as well as non-parametric kernels via random Fourier features and multilayer perceptrons. Model selection across families uses BIC and cross-validation. On synthetic data the framework recovers the ground-truth kernel across regimes, and BIC reliably selects the generating family with margins of thousands of nats. We then apply the framework to six empirical datasets spanning congressional roll-call votes, controlled numerical-estimation experiments [3], online prediction markets, VK social-network traces [4], and Reddit discussions (Fig. 1b). The result is consistent: BIC prefers either DeGroot or a near-flat sigmoidal bounded- confidence kernel in nearly every case, with congressional data the lone exception where sharp bounded- confidence dynamics dominate. Repulsive kernels are never selected. Non-parametric kernels corroborate the parametric story, showing only weak dependence on opinion distance. Microscopic kernel inference from observational social data is signal-starved: what survives is consistent evidence for assimilative social influence across domains, and a quantitative framework for asking the question elsewhere

ABSTRACT. Healthcare systems are complex adaptive systems that can self-organize when stressed, such as during periods of significant patient surge. Here, we look at patient surge in the critical care environment during the COVID-19 pandemic. Bioinspired computing methods including Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO) are utilized to load balance critical care (ICU) beds across New York State (NYS). First, GA and PSO are explored to load balance ICU beds to reduce patient mortality and optimize ICU resource allocation. Next, ACO is applied to identify optimal patient transfer routes between overloaded and underloaded facilities, decompressing overwhelmed facilities and regions. The simulation of load balancing and patient transfers utilized real data provided by the NYS Health Electronic Data Response System (HERDS) to dynamically adapt to the evolving pandemic. Results show that an adaptive PSO model successfully improved ICU capacity across NYS leading to improved ICU bed allocation and reduced fatalities. ACO effectively and efficiently decompressed overloaded facilities and regions while minimizing transfer costs. This resulted in a 90% reduction in overloaded hospitals and a decrease in average ICU load from 99% to 67%. Adaptive, decentralized, self-organizing systems can effectively coordinate resource allocation through load balancing and patient transfers, leading to increased resiliency in the event of patient surge. The results support adoption of real-time deployment of the adaptive PSO and ACO model as part of a dynamic response to future healthcare surges such as pandemics or mass-casualty events.

ABSTRACT. How does a narrative mutate as it spreads? Social contagion research has extensively modeled how information moves from person to person, but the parallel question, how the story itself changes under diffusion, has received less formal treatment. We introduce narrative contagion as a second-order diffusion process in which the mechanism of transmission is itself subject to selection pressure. Drawing on cultural epidemiology and meme theory, we treat each narrative variant as a culturally transmitted unit that can be copied, recombined, and modified as it passes between actors, gaining or losing transmission strength under selection as it goes. The 1692 Salem witch trials provide exceptional historic data for this purpose: accusers were required to specify not merely that harm occurred, but how, producing a structured record of supernatural mechanism claims across 983 trial documents. We manually annotate these as directed relational events (Accuser → Mechanism → Accused) across eight mechanism categories, constructing a temporal bipartite network that embeds narrative content directly in its edge structure. Using network-based diffusion analysis, we test whether mechanism adoption follows network proximity faster than a null model of independent invention, and whether mechanism diversity, measured by Shannon entropy, first expands then contracts as dominant motifs crowd out alternatives. Beyond Salem, this pipeline offers a replicable framework for studying narrative escalation in collective belief episodes more broadly.

ABSTRACT. Scientific knowledge often develops within separate disciplinary traditions. As a result, causal mechanisms that are well established in one field may remain invisible to another field facing structurally similar problems but using different concepts, vocabularies, and empirical traditions. This problem is especially important for interdisciplinary research, where mechanisms may recur across domains but remain disconnected in literature. This paper introduces CausalBridge, a multi-agent LLM platform for discovering hidden causal links across scientific domains. CausalBridge is designed as an end-to-end computational discovery platform. Specialized LLM agents ingest scientific literature, extract causal claims, canonicalize constructs, construct disciplinary causal knowledge networks, identify cross-domain anchors, generate candidate hypotheses, validate novelty, and rank outputs for human review. Rather than relying on keyword search or single-prompt reasoning, CausalBridge decomposes interdisciplinary discovery into coordinated and inspectable agentic modules. This allows the platform to identify causal bridges that may be structurally or mechanistically meaningful even when the relevant literature do not share terminology.

At the core of CausalBridge is an anchor-walk framework over scientific causal networks. Each discipline is represented as a directed causal knowledge network whose nodes are canonicalized constructs and whose edges are literature-attested causal relationships annotated by polarity, mechanism type, and frequency of explicit assertion. I demonstrate the platform using criminology and natural-disaster research, two socially consequential domains with potential mechanistic overlap but distinct disciplinary vocabularies. The resulting networks contain 1,300 nodes and 4,006 edges for criminology, and 1,264 nodes and 4,319 edges for natural disasters. Applied to this domain pair, CausalBridge generated 2,521 candidate hypotheses across multiple discovery paths, drawing on approximately 6,000 unique papers and 18,700 abstract-level stance classifications from OpenAlex and Semantic Scholar. The validation module assessed evidence quantity, citation- and recency-weighted quality, conclusion stance, conceptual coherence, vocabulary-mismatch redundancy, and grey-literature coverage. After filtering, the platform retained 327 radical-novelty candidates and 21 bridging candidates, alongside 238 refinement and 1,563 confirmatory hypotheses.

Representative radical-novelty candidates connect social-contagion mechanisms in criminology to disaster-response failure modes, revealing structurally supported causal possibilities not explicitly developed in either literature. By combining literature-scale causal extraction, complex-network alignment, agentic hypothesis generation, and multi-stage novelty validation, the platform transforms fragmented scientific literature into a navigable space of causal analogies, missing mechanisms, and testable cross-domain hypotheses.

ABSTRACT. Climate change has raised major concerns about its impacts on agricultural yields and irrigation water demand. Growing-season planning has become increasingly challenging due to highly variable environmental conditions. Integrating climate, crop, and water-management modules creates a complex system that farmers and local decision-makers must navigate to achieve resource-efficient planning. The combined impacts of projected environmental factors on crop yield and irrigation demand are difficult to predict and evaluate. However, few modeling frameworks have been developed to assess climate-responsive adaptation strategies for an optimized irrigation management under near- and mid-term climate scenarios.

This study develops a high-resolution, climate-responsive optimization module within an integrated climate-crop framework for data-driven irrigation planning. The framework is designed to couple downscaled climate projections with a process-based crop model and adaptive irrigation decision rules. Our modified version of the SIMPLE crop model (SIMPLE+) uses daily water-stress-factor thresholds to estimate irrigation requirements for different crops throughout the growing season. The case study focuses on the Navajo Nation, a drought-prone Indigenous region facing water infrastructure constraints and food insecurity, where water stress is a dominant growth-limiting factor and rainfed systems achieve less than 10% of potential yield for major crops. These conditions highlight the importance of adaptive irrigation management in arid regions.

The optimization framework evaluates tradeoffs between crop yield and irrigation water demand across climate scenarios. Different local and regional optimization objectives are explored by treating the daily water-stress threshold as a decision variable. The coupled climate-crop framework enables adaptive irrigation planning under uncertain future climate conditions by dynamically linking projected environmental conditions with crop water-demand dynamics. The framework enables localized irrigation planning and water allocation and is scalable to other regions for climate-resilient agricultural water-resource management.

ABSTRACT. Violence in armed conflict does not occur randomly. Drawing on complex systems frameworks of spatial self-organization, temporal burstiness, and emergent feedback, I examined how the structure of conflict shapes civilian collective action in sub-Saharan Africa. I analyzed 5.9 million location-month observations across 17,524 unique locations in 51 countries (1997–2024) using georeferenced event data from the Armed Conflict Location and Event Data Project (ACLED). Spatial-temporal clustering analysis identifies 1,280 conflict hotspot locations (7.3%) exhibiting persistent, intense, or volatile activity. Violence within these locations displays strong temporal autocorrelation (β = 0.161, p < .001), consistent with path-dependent entrenchment. Burstiness analysis [B = (σ − μ)/(σ + μ); [1] reveals high temporal clustering in both protest events (B ≈ 0.81) and violence against civilians (B ≈ 0.83). Hotspot locations exhibit significantly lower burstiness than non-hotspot locations (p < .001), indicating sustained rather than episodic dynamics in entrenched conflict zones. Temporal feedback analysis reveals an asymmetric Granger-causal structure: protest burstiness predicts subsequent violence burstiness (χ2 = 23.85, p < .001), but violence burstiness does not predict protest burstiness (χ2 = 4.18, p = .124). This unidirectional ordering implies that civilian mobilization shapes how violence clusters in time, while armed-group violence does not reciprocally organize civilian protest timing—an asymmetry with direct implications for early warning system design. Location fixed-effects analysis reveals substantial spatial heterogeneity: the protest-violence association is approximately five times stronger in conflict hotspots than in non-hotspot locations (interaction β = 0.213, p < .001; Figure 1). These findings provide substantial support for a complex systems account of civilian-armed actor dynamics: violence self-organizes into identifiable spatial clusters, both actors exhibit bursty temporal behavior consistent with coordinated human dynamics [2], and protest burstiness feeds forward into violence timing through a unidirectional causal channel. I interpreted the differential spatial effect as most plausibly reflecting armed group territorial entrenchment, though direct mechanism identification requires future research.

ABSTRACT. Open-ended evolution does not require genes, replicators, or an externally specified fitness function: stochastic assembly combined with differential persistence is sufficient to generate sustained novelty, scaffold-level dominance, and structured population-level order in open non-equilibrium systems. We introduce Stability-Driven Assembly (SDA) [1], in which stochastic interactions generate patterns whose persistence depends on their stability. Longer-lived patterns accumulate and are sampled more often, closing a feedback loop: persistence shapes composition, composition biases sampling, and fitness-proportional selection emerges endogenously from differential persistence alone.

The dynamics are path-dependent: new structures emerge that did not exist at the outset, and their persistence reshapes the conditions for what can emerge next. Standard system dynamics formalisms hold the state space fixed and cannot capture this. Chemical simulations on SMILES fragments [2] show scaffold dominance, sustained novelty, and heavy-tailed rank-abundance distributions over thousands of generations.

The same persistence-weighted feedback governs chemical evolution, organizational evolution in industry ecosystems, and institutional change. In this view, replicators are an efficient implementation of a more general population-memory function, not its foundation. Evolution shifts from being a biological capability that requires replicating genes to being a feature of open dynamical systems with differential persistence and population feedback.

References:

[1] Adler, D. (2025). Stability-Driven Assembly Theory. Journal of Theoretical Biology. https://doi.org/10.1016/j.jtbi.2025.112352

[2] Adler, D. (2026). How Information Evolves: Stability-Driven Assembly and the Emergence of a Natural Genetic Algorithm. https://arxiv.org/abs/2601.17061

ABSTRACT. Many consequential life outcomes depend not only on the option chosen, but on how coherently an individual’s preferences are structured at the moment of commitment. Under constrained choice, early commitment, and limited reversibility, observed choices need not reflect a well-integrated preference structure. We define preference alignment as a pre-commitment, relational property of choice sets: the degree to which an individual’s stated options cluster in a network of alternatives, rather than spanning unrelated ones. This is a horizontal dimension of decision quality, distinct from the ex-post vertical/horizontal mismatch typically studied in labour and education economics. Higher education offers a clean empirical setting: applicants reveal ordered preferences in centralized admission systems before cutoff-based assignment. From population-level data (Chile, 2012–2019), we build a co-application network of degree programmes; edges link pairs with above-chance joint listing (φ-correlation, P ≤ 0.05), so proximity captures shared academic content, career trajectories, and institutional features. For each applicant k with portfolio P_k of size N_k, the preference distance C_k is the average shortest-path distance d_ij over all programme pairs (i,j) in P_k. Low C_k flags an aligned portfolio; high C_k a dispersed one. We estimate (i) descriptive associations between C_k and first-year retention and (ii) a fuzzy regression discontinuity design (RDD) at programme-specific admission cutoffs, with heterogeneity by C_k. Using administrative data from over 1.2 million applicants in Chile and a replication in Portugal, we find three robust patterns (Fig. 1). First, first-year retention declines monotonically with C_k across all enrolled-score tertiles (Fig. 1A); aligned portfolios sustain above 80% retention regardless of score, while the most dispersed portfolios converge near 40%. Second, the fuzzy RDD validates a discrete first-stage jump in first-choice enrolment at the cutoff (Fig. 1C) and a positive LATE on retention (LATE = 0.308, SE 0.030, P < 0.01; Fig. 1D). Third, and critically, the LATE interacts with alignment (βTreatment×Distance = 0.074, SE 0.007, P < 0.01): the retention gain from first-choice enrolment is concentrated where preferences are aligned and attenuates sharply as C_k rises. The interaction replicates in Portugal in sign and magnitude, indicating a behavioural regularity rather than a context-specific artefact. In predictive models, preference distance ranks among the three most informative features of early persistence, surpassing several standard academic and socio-economic covariates. Preference alignment thus identifies a relational, network-structured dimension of decision quality: institutional assignment translates into sustained engagement only when it is anchored in an internally coherent choice set.

ABSTRACT. Teachers hold a prominent place in modern societies, particularly where education is compulsory and widely institutionalized. This ubiquity obscures an underlying puzzle: Why do societies assign some individuals solely to the instruction of others? This question is especially enigmatic for dedicated teachers, who invest their labor in cultivating others’ skills but do not themselves participate in the productive activities for which their students are being trained. To address this question, we develop a simple, mathematically tractable model of teaching and learning in a population that shares a common productive aim. We identify a tradeoff between the size of the workforce and its collective level of expertise; and we analyze the optimal proportion of a population that should serve as teachers, across diverse scenarios. We find that a population must exceed a critical size before it is beneficial to allocate anyone as a dedicated teacher at all. Subsequently, the peak demand for teachers is achieved at an intermediate population size and never exceeds one half of the population. For more complicated tasks, our analysis predicts the optimal allocation of teachers to instruction across different levels of expertise. Although our model is not evolutionary per se, this population-level account lays a foundation for understanding the adaptive advantage of dedicated teachers in both human and nonhuman populations.

ABSTRACT. We implement the reduced transfer entropy (TE) estimator described by Kirkley \cite{Kirkley2025}, a finite-data estimator that uses the minimum description length (MDL) principle for automatic significance selection, to electrophysiology recordings from the ventral and dorsal CA1 regions of the hippocampus in both healthy and amyloid pathological mice (APP/PS1). From the MDL-thresholded TE networks we compute weighted directed reciprocity and find region- and pathology-dependent differences: APP/PS1 shows, on average, higher reciprocity than control in vCA1 but lower reciprocity in dCA1. We also compare TE distributions across regions and examine how reduced TE and reciprocity vary with inter-neuron distance, revealing distinct distance-dependence profiles between regions and pathologies, as can be seen in Figure \ref{fig:main}. These results indicate a potentially useful perspective for examining how amyloid pathology in a model of Alzheimer's may differently affect the spread of information in neighboring brain regions.

ABSTRACT. Disaster-related information is shared on social media (decision) and spreads through social networks, shaping disaster awareness, preparedness, response, and recovery for resilient disaster management. Given different disaster characteristics, these communication dynamics are likely to vary across disaster types [1]. However, we still lack knowledge on how different disasters shape social media communication. Existing works have primarily focused on information diffusion using Susceptible-Infected-Recovered (SIR) models. Yet, communication dynamics involve individual decisions to share posts, as well as diffusion. To address the gaps, we construct social networks from X data using keywords of 2012 United States (U.S.) droughts and wildfires. We then develop an integrated modeling framework that couples multi-agent reinforcement learning (MARL) for sharing decisions and SIR models for information spread. In the drought communication network, U.S. news media at both local and national levels serve as central information hubs. Individual influencers, including journalists and researchers, are also play a key role for drought communications along with the new media. Drought-related communication forms a regional cluster across Nebraska, Iowa, Colorado, Kansas, and Oklahoma, while wildfire communication exhibits a dominant connection between Idaho and Colorado (Figs. 1A-1B). Drought information spreads relatively gradually, whereas wildfire information spreads abruptly, driven by the nature of slow-onset and abrupt disasters (Figs. 1C-1D). These behaviors are effectively captured by the SIR-MARL framework. In particular, the model reflects systematic, adaptive communication behaviors, compared with SIR-only models. This work reveals how drought and wilㅇfire communications are characterized heterogeneously, ultimately guiding policymakers to establish disaster-specific and effective disaster planning.

ABSTRACT. Politically divided societies are also often divided emotionally: people like and trust those with similar political views (in-group favoritism) while disliking and distrusting those with different views (out-group animosity). This phenomenon, called affective polarization, influences individual decisions, including seemingly apolitical choices such as whether to wear a mask or what car to buy. We present a dynamical model of decision-making in an affectively polarized society, identifying three potential global outcomes separated by a sharp boundary in the parameter space: consensus, partisan polarization, and nonpartisan polarization. Analysis reveals that larger out-group animosity compared to in-group favoritism, i.e. more hate than love, is sufficient for polarization, while larger in-group favoritism compared to out-group animosity, i.e. more love than hate, is necessary for consensus. We also show that, counterintuitively, increasing cross-party connections facilitates polarization, and that by emphasizing partisan differences, mass media creates self-fulfilling prophecies that lead to polarization. Affective polarization also creates tipping points in the opinion landscape where one group suddenly reverses their trends. Our findings aid in understanding and addressing the cascading effects of affective polarization, offering insights for strategies to mitigate polarization. This analytically tractable model reconciles seemingly contradictory findings in the literature and provides a theoretical foundation to study and mitigate harmful polarization dynamics on social networks.

ABSTRACT. Signed networks capture positive (friendship) and negative (antagonistic) relationships in social networks. Structural balance and status theory are the dominant frameworks for studying these ties, yet they fail to explain imbalanced triads and overlook competing mechanisms such as homophily, reciprocity, and popularity. To disentangle these mechanisms, we combine over 70 node and edge features (see Figure 1(a)) in a logistic regression classifier for edge sign prediction on a temporal high school network of 888 students [1]. We apply inductive learning, partitioning the network by course and training on one course while testing on others to avoid data leakage. Our model achieves high balanced accuracy, AUC, and F1-scores across held-out courses and snapshots (Figure 1(b)). Feature importance analysis reveals consistent predictors: reciprocal tie strength, friend influence, and positive out-degree for positive edges; negative out-degree and negative reciprocal links for negative edges (Figure 1(c)). Surprisingly, certain triad-based features (BBpp, FFpp) predict signs that violate previous social theories, suggesting additional mechanisms drive signed tie formation. These findings offer guidance for improving the explainability of deep learning models for link sign prediction, which typically rely solely on balance and status theory.

ABSTRACT. The social network surrounding child development is undergoing a sequence of structural transformations whose consequences remain poorly characterised. We argue that network topology offers a productive lens for rising adolescent mental-health and social difficulties, reframing what are typically treated as separate cultural, technological, and clinical concerns as facets of a single shift in children's primary social environments.

We distinguish four idealised topologies that have successively dominated those environments (Figure 1): (i) dense intergenerational kin networks, with high clustering and multiple emotionally available adults; (ii) homophilous peer clusters, in which primary attachment shifts toward same-age peers and intergenerational bridging weakens, the structural pattern underlying Neufeld and Maté's account of "peer orientation"; (iii) scale-free parasocial structures, in which children orient toward a small set of high-degree influencer nodes whose connectivity is asymmetric; and (iv) isolated human–AI dyads, in which a child's most emotionally responsive interlocutor is a chatbot or AI companion with no onward connectivity to the wider social graph.

Each transition removes structural features on which the previous environment relied. Behaviours and norms typically spread as complex contagions requiring reinforcement from multiple network neighbours (Centola); redundant ties, triadic closure, and intergenerational bridges are therefore the substrate on which protective norms, emotional regulation, and corrective social feedback propagate. Healthy mood spreads across adolescent friendship networks as a complex contagion, with sufficient healthy-mood ties roughly halving the probability of developing depression over a 6–12 month window (Hill et al., 2015); recent registry evidence further suggests that mood, anxiety, and eating disorders may transmit through adolescent peer networks.

These successive shifts plausibly constitute social-network analogues of the critical transitions described in the broader complex-systems literature, and they are likely to shape children's socialisation and psychological development in ways that linear cause-effect studies cannot detect. By treating structural properties as primary, the network perspective surfaces effects that are emergent rather than additive and identifies where in the social fabric intervention is likely to be tractable. We offer this lens as a complement, not a substitute, for variable-centred work: each captures something the other cannot.

ABSTRACT. Healthy disagreement among peers is vital to the production of knowledge. Disagreement is also social, influenced by a work's visibility, topic, and social concerns of reputation and reprisal. We present an investigation into the social structure of debate in science. We examine two datasets representing severe and mild disagreement, respectively. First is 2,000 critical letters that appear in APS journals. Second is 23,000 "contrasting" (and 350,000 "supporting") citations sourced from Scite.ai. Disagreement in science tends to be local. "Contrasting" citations are more often issued within the same venue and field, and to more topically similar papers than neutral ones (Fig. 1a). Notably, the most highly-cited (visible) papers are less likely than expected to receive a contrasting citation, suggesting that visibility is not the sole driver of disagreement (Fig. 1b). Social forces shape disagreement. Senior authors are disproportionately involved compared to other authors published in the same journal (Fig. 1c), and are more often men and from Europe (results not shown). Focusing on the case of APS, the "social distance" spanned by authors writing and receiving critical letters is greater than the baseline of neutral citations (Fig. 1d). Disagreement is structurally distinct from neutral citation. We build a journal citation network and contrast the weighted degree and eigenvector centrality of journals using only edges of each citation type. "Contrasting" citations are more dissimilar to mentioning citations than are supporting across both node centrality metrics (Fig. 1e). Collectively, these results suggest that disagreement in science is directed locally against similar papers, but not against direct or near collaborators. Further, disagreement is a structurally distinct form of scholarly communication.

ABSTRACT. The evolution of product-level sales can be understood as the output of a complex system in which heterogeneous consumers make purchasing decisions based on personal preferences, macroeconomic conditions and companies’ sales strategies. This interaction becomes particularly intricate in consumer electronics markets, where multiple product generations coexist and compete over time. In this context, reliable demand forecasting is essential for operational planning, enabling firms to satisfy demand while reducing inventory costs. Although machine learning methods have become increasingly prevalent due to their predictive performance, statistical modeling remains competitive, offering interpretability and requiring less extensive training data. In this work, developed in collaboration with Sony Corporation, we extend a statistical framework for forecasting weekly TV sales in the United States using data from 2014 to 2025. The weekly resolution introduces irregular seasonality, which we address through a decomposition-based method using the concept of analogous weeks. In order to predict sales of individual TV models, we use a top-down approach by independently forecasting the total TV sales and the shares of individual products. For share prediction, we introduce a model-level representation of product lifecycles using smooth-trapezoid functions, enabling the characterization of the long-term sales behavior including growth, maturity and decline phases. Short-term fluctuations are influenced by price dynamics, which are incorporated into the method using both own- and cross-price elasticities. We also include autoregressive corrections to account for residual temporal correlations. Backtesting results for 42 relevant TV models for the year 2025 show that the proposed method achieves levels of accuracy comparable to those obtained by expert human forecasts, with around 20% average monthly error. As such, this approach provides a practical tool for business applications that may replace or be integrated with judgmental forecasting. Retaining interpretability, it also serves as a framework for understanding the mechanisms underlying sales dynamics, highlighting the role of product lifecycle phases and pricing strategies.

ABSTRACT. Technological progress is rarely about improving a single feature. From efficiency and performance to cost and stability, advances unfold across multiple, often competing objectives, giving rise to trade-offs that define a moving Pareto frontier of what is jointly attainable. While Pareto frontiers capture these trade-offs in static form, little is known about how these frontiers evolve as innovation progresses. Tracing the state-of-the-art trajectories in three domains of rapid innovation—large-language models, perovskite solar cells, and computer hardware—we find that progress follows a universal two-phase pattern. Early progress follows a coordinated phase, in which advances simultaneously improve all objectives and the frontier converges to a single point. Beyond a critical threshold, coordination collapses, giving rise to a fragmented phase with intensified trade-offs and objective-specific leaders. We explain this pattern with a minimal stochastic model of innovation. Two forces compete: (i) a steady pressure toward specialization from moves that help one objective while hurting the other, and (ii) a finite reservoir of “win–win” moves that advance all objectives but become exhausted as the system improves. As universally beneficial moves diminish, the system tips from coordination to fragmentation in a sharp dynamical phase transition whose timing scales with problem complexity —roughly the number of design dimensions times its logarithm— and location shifts with the number of inherent conflicts between objectives. The model yields a simple scaling collapse that unifies trajectories across technologies and matches the observed transitions. Spontaneous divergence thus emerges as a general organizing principle of technological progress, offering testable forecasts for when frontiers will bifurcate, cautioning against over-extrapolating from early unified leaders, and informing the design of R&D portfolios, standards, and policy in domains where multiple goals must advance together.

ABSTRACT. High-throughput technologies enable the reconstruction of large-scale biological networks, such as gene regulatory networks. However, the increasing size and complexity of models pose challenges for computational analysis and network inference. Nevertheless, finite experimentally-validated automata models enable dynamical simulations and analysis of biological regulation revealing key signaling mechanisms and identifying potential therapeutic targets [2]. They consider its components (drugs, genes, proteins, etc) as Boolean automata, while causal regulatory interactions drive the dynamics of micro-level node states (eg.: gene expression) and macro-level network configurations (eg.: cellular state) [1]. This Boolean network (BN) framework provide a robust, explainable, and computationally tractable approach that captures both network structure and dynamics. In this study we showcase how the conditional effective graph (CEG) representation, defined below, of finite automata models can guide us to inferences in large-scale networks.

Representing how each automata node and regulatory interaction contributes to BN dynamics logic can be done via the effective graph (EG) [3], which is obtained by removing logical redundancy via Boolean minimization [4].It quantifies the probabilistic parameter edge effectivenes eji ∈ [0,1] weight, and its aggregate node-level effective connectivity, ke(xi) = ∑_j^ki(eji) [3]. Furthermore, by conditioning a subset of variables to a specific state (ON or OFF), the effective connectivity of the EG is altered leading to the conditional effective graph (CEG), which was also proposed for analysis of the EG with partial information [3]. This way, the CEG enables direct analysis of network dynamics without explicit enumeration of the full dynamical landscape or Monte Carlo simulations. As logical constraints propagate through the network, some nodes and edges can become entirely redundant or constant. Conversely, other nodes and edges remain dynamically viable, meaning they can still take either state and remain available for intervention. To quantify how much of a BN remains dynamically viable under a given conditioning, we define actionability at node and edge levels. Node-level (edge-level) actionability is defined as the proportion of nodes (edges) that remain viable, not redundant nor constant. We introduce new methods and guidelines in the open-source CANA Python package to compute actionability and visualize the EG and CEG of any BN [5].

To demonstrate the utility of the CEG, we apply our framework to the BT474 short-term ErbB signaling network model [6] and visualize how partial state conditioning affects signal propagation within the network (Figure 1, A). Anewcorresponding tutorial demonstrating this analysis is available as a Python notebook in the CANA repository [7]. We also compute the ground-truth controllability of 41 experimentally-validated BN models of biochemical regulation from the literature using the attractor control introduced by Gates and Rocha [8], to identify the set of true driver nodes of each network (nodes that collectively are sufficient to control dynamics to any attractor). To evaluate the merit of the actionability measure, we condition every network node as ON or OFF and compare action ability for driver and non-driver nodes (Figure 1. B). We observe statistically significant differences in actionability, proving that conditioning driver nodes (individually) reduces actionability more than conditioning non-driver nodes, consistent with their role in BN controllability. In future work, we will extend the analyses to substantially larger BNmodels (eg.: ∼ 1000 nodes), leveraging the scalability of the CEG framework.

ABSTRACT. Time series forecasting (TSF) methods are typically evaluated on benchmark datasets with straightforward and periodic dynamical trends (e.g., electricity consumption and traffic occupancy), where uncertainty appears as local fluctuations around a deterministic mean. This differs from intrinsically stochastic systems, such as biomolecular systems, where stochasticity is not merely noise but a fundamental driver of function. At the molecular scale, thermal fluctuations, solvent interactions, and conformational heterogeneity give rise to a landscape of metastable states, where distributions of conformations and their transition probabilities,``not single trajectories", encode activity. In order to enable TSF of biomolecular systems, we argue for a new class of benchmarks for TSF based on stochastic dynamics, where success requires learning the distribution of possible futures rather than only predicting one realized trajectory. These benchmarks include particles evolving under deterministic forces with thermal fluctuations, as well as molecular dynamics trajectories of alanine dipeptide. In these, we find that a simple normalizing flow-based probabilistic method, which we call NFTSF, outperforms state of the art probabilistic TSF methods on probability matching metrics by up to a factor of two, while reducing sampling time in short horizon by approximately a factor of seven. These findings show that the current probabilistic TSF evaluation was not designed for systems in which stochasticity itself is the object to be learned.

ABSTRACT. Boolean networks (BNs) are widely used to model biochemical regulation, but methods that require enumeration of their state space are not feasible for even modestly sized networks [1]. Methods based only on the Interaction graph (IG) of BN (AKA structure-only methods), offer scalable alternatives but miss key dynamical information by treating all interactions equally, resulting in incorrectly predicted dynamics and controllability [2]. A scalable alternative is the effective graph (EG), obtained by removing logical redundancy from Boolean Automata [3]. This parsimonious weighted graph representation integrates (higher-order) dynamical information into probabilistic parameters that quantify how effective each interaction is in driving downstream dynamics. We show that analysis of EGs allows us to infer three dynamical properties of BN:

Reachability characterizes how signals propagate in a BN from a given node to other nodes. It can be efficiently estimated by computing path length in the EG via the product of edge effectiveness weights, significantly outperforming IG-based predictions [3]. We show that the reachability prediction is significantly improved if the effectiveness weights of the EG are refined to account for input biases, accounting for unequal probabilities of "ON" or "OFF" signals into a given node.

Controllability of BN is known to be constrained by the Feedback vertex set (FVS) of the IG [4]. That is, the set of sufficient driver nodes for pinning control of the ensemble of BNs defined by the same IG, is the minimum set of nodes that break all its loops. Thus, for a specific BN, the FVS is a superset of the true driver set, which can be smaller [3]. Furthermore, FVS does not distinguish between loops that are more or less effective in propagating signals. We show that the effectiveness strength of loops in the EG provides better estimates of the fraction of controlled state space compared to the IG. Furthermore, the conditional effective graph (CEG) [3] provides an even more accurate estimate, by accounting for the pinned nodes' states (see figure).

Dynamical modularity refers to subgraphs that tend to capture or trap signal propagation or long-term dynamics, which we study via weakly and strongly connected components (WCCs and SCCs) of the EG under effectiveness thresholding. In the experimentally-validated Cell Collective models [5], WCCs remain intact even at thresholds up to [0.4, 0.6], while SCCs disintegrate at lower thresholds, suggesting that signal propagation in biochemical networks is robust while attractor control is more fragile.

ABSTRACT. Living and artificial systems must sustain robustness in complex, chaotic environments, motivating the need for a foundational information thermodynamic metric for complex systems \cite{lewinComplexityLifeEdge1999, karacaMultiChaosFractalMultiFractional2022}. Such systems combine short-range dynamics, characterized by exponential growth in state space, $\lim\limits_{N \rightarrow 0}W(N) \sim \exp N^\alpha$, where $\alpha$, is the stretch exponential potential and long-range dynamics, characterized by power law growth in the state space, $\lim\limits_{N \rightarrow \infty} W(N) \sim N^{d+\frac{1}{\kappa}}$, where $d$ is the dimensionality of the system and $\kappa$ is the nonlinear deviation from linear systems. All of the universality classes for complex systems identified by Hanel-Thurner to satisfy the first three Shannon-Khinchin (SK) axioms \cite{hanelComprehensiveClassificationComplex2011}, are a combination of these short- and long-range interactions. The state space growth of all complex systems can be approximated as a smooth interpolation of the short and long range dynamics, $W(N) \approx \exp_{\alpha \kappa}^{1+d\kappa} \left(\frac{N^\alpha}{\alpha N_\sigma}\right)$, where $\exp_a x \equiv (1+a x)\frac{1}{a}$, mediated by an informational scale, $N_\sigma$. From this foundational model of the interactions in a complex system, the coupled entropy, is proven to be the unique, stable, and universal generalization of entropy \cite{nelsonUniquenessCoupledEntropy2025}.

ABSTRACT. Fault tree analysis (FTA) is a systematic method for identifying combinations of component failures that lead to system-level failures, widely used in complex systems modeling for high-risk reliability engineering (e.g. rockets, nuclear power plants, etc.). While existing R packages provide FTA functionality, they lack integration with the broader tidyverse ecosystem, making it difficult to programmatically construct fault trees, visualize them using modern graphics tools, and integrate FTA workflows into reproducible data analysis pipelines.

This paper introduces tidyfault, an R package that provides a unified, pipe-friendly workflow for fault tree analysis, built on R’s ‘tidyverse’ principles. By representing complex system logic as "tidy" data frames of nodes and edges, the package allows researchers to leverage the full power of functional programming and visualization in R. The package implements a pipeline for processing fault trees from nodes and edges through boolean equation extraction, truth table generation, and minimal cutset identification using a custom R implementation of the Method of Obtaining Cut Sets (MOCUS) algorithm. Additional functions support binary scenario evaluation, probability computation, and ggplot2-compatible visualization, helping bring FTA fully into modern R data science workflows. Unlike traditional GUI-based FTA tools, tidyfault allows for the batch evaluation of thousands of failure scenarios, making it uniquely suited for the high-dimensional simulations typical of complex systems research.

We demonstrate the package’s functionality through case studies involving security systems and database failures. These examples highlight how tidyfault can identify critical vulnerabilities and single points of failure within socio-technical systems. It is the authors’ hope that tidyfault’s reproducible, programmatic fault tree analysis framework will help a new generation of systems engineers scale up their FTA analyses to match the growing size and scope of modern complex systems.

ABSTRACT. The expansion of bicycle-sharing systems in large cities provides an interesting case study for analyzing urban mobility from a complex systems perspective. In this work, we analyze the evolution of the public bicycle network in the City of Buenos Aires, Argentina between 2010 and 2025 (Fig. 1). We model the system as a dynamic network where stations are nodes and trips define weighted edges, and study its structural properties as the system grows. Using longitudinal data on stations, users, and trips, we quantify the increase in infrastructure, usage, and connectivity, highlighting the growth in the spatial coverage, connectivity and trip diversity.

From a mobility standpoint, we identify clear temporal and spatial patterns. Weekday usage reflects strong pendular behavior, with flows toward the Financial District as a primary work-related attractor, while weekend patterns emphasize recreational areas such as Palermo. We also examine the interaction with other transportation modes, with key connections to major train stations such as Constitución and Retiro. Differences across weekdays and weekends reveal distinct usage regimes in both trip volume and user composition.

At the system level, the network becomes denser and more interconnected over time, going from 4 stations and 4 links per station in 2010 to 407 stations and 171 links per station in 2025. The network grows from the central business district, reaching northern neighborhoods first, while the south of the city remains less connected. Finally, we explore the impact of the COVID-19 pandemic as a potential shift in mobility patterns. These results illustrate the role of bicycle-sharing systems in shaping urban mobility.

ABSTRACT. A core question in complex systems is the problem understanding the nature of high-dimensional, many-body interactions - and over the decades, proposals have linked these structures to everything from “emergence” to “consciousness.” This is particularly relevant to neuroscience, as the brain is a paradigmatic example of a complex system showing complex behaviors emerging from the interaction of billions of individual neurons. The study of genuine higher-order interactions has been historically hampered, however, by mathematical and theoretical difficulties in studying such structures directly. Instead, a common approach has been to study them obliquely, either by building multi-partite interactions out of collections of lower-order dependencies (as in functional connectivity network analysis) or reducing the overwhelming dimensionality down to a tractable form with manifold-learning algorithms like PCA. In this talk, I argue that these approaches systematically fail to accurately represent genuine higher-order interactions. Using the formal language of multivariate information theory and formalizing “genuine higher-order interactions” with synergistic information, this talk presents theoretical and empirical evidence that both statistical networks and dimensionality reduction-based approaches to modeling complex systems are systematically blind to synergistic higher-order interactions. In neuroimaging data, we find that, while synergistic information is quite common, network and PCA-based approaches consistently fail to find synergy-dominated sets of elements, instead being sensitive to lower-order redundancies. These higher-order synergies form a kind of “dark information”: potentially accounting for a huge part of all the “structure” in a system, but largely invisible to established, mainstream analytical methods. Finally, we end with a discussion of why synergistic information should be of widespread practical interest, showing links to causal colliders and evidence that synergy is associated with complex cognitive processes over the course of development.

ABSTRACT. Many economic models implicitly assume continuity across state space despite the existence of hard lower bounds such as insolvency, unemployment, or zero purchasing power. We argue that these zero-crossings generate absorbing and near-absorbing boundaries that qualitatively alter the dynamics of economic systems.

We consider a heterogeneous-agent adaptive system in which agents participate in market exchange only when their effective purchasing power remains positive. Let agent wealth evolve according to: \[ w_i(t+1) = w_i(t) + y_i(t) - c_i(t) - r\,d_i(t), \] where $w_i$ denotes wealth, $y_i$ income, $c_i$ consumption, and $d_i$ debt obligations. Participation in market signalling is governed by: \[ p_i(t) = \begin{cases} 1 & \text{if } w_i(t) > 0 \\ 0 & \text{if } w_i(t) \le 0. \end{cases} \]

Aggregate demand is therefore $D(t) = \sum_i p_i(t)\cdot\mathrm{need}_i(t)$, rather than the total underlying need in the system. As agents cross the zero boundary, unmet need ceases to be encoded into market demand. The economy undergoes discontinuous changes in information-processing structure: agents may continue to possess needs, but no longer transmit economically effective signals.

Sufficiently large concentrations of zero-crossings generate a loss of adaptive capacity at the system level. As the active population $A(t) = \sum_i p_i(t)$ contracts, exchange, coordination, and recovery dynamics weaken nonlinearly. The resulting dynamics produce absorbing and near-absorbing regions in state space in which re-entry into productive participation becomes increasingly unlikely.

At the aggregate level, the economy evolves as $\dot{x} = F(x,\, P(w),\, \mathcal{N})$, where $P(w)$ denotes the evolving wealth distribution and $\mathcal{N}$ the interaction network. Zero-crossings deform basin geometry by altering participation structure and signal propagation across the adaptive network. Under sufficiently high exclusion, desirable macroeconomic states cease to remain dynamically reachable under feasible interventions.

This framework connects economic fragility to absorbing-state transitions, network fragmentation, and critical loss of functionality across complex systems. Systemic resilience requires preserving both local stability and the reachability of desirable states through maintained participation and information flow.

\enlargethispage{3\baselineskip}

%%====================================================================

\vspace{2pt} \begin{center} \includegraphics[width=\textwidth,height=1.3in,keepaspectratio]{fig_A_three_failure_modes.png}

\vspace{2pt} {\footnotesize\itshape \textbf{Figure 1.} Signal collapse (A), absorbing-state hysteresis (B), and risk amplification (C) at the zero crossing; floor $\varepsilon>0$ bounds all three.} \end{center}

ABSTRACT. Indonesians, and especially low and middle-income (LMI) Indonesian families rely heavily on 2-wheeler vehicles (henceforth, 2WVs including motorbikes, motorcycles, etc.) as the main mode of transport. For these LMI families, purchasing passenger cars is cost-prohibitive and non-urban public transit networks are often underdeveloped and have prohibitively long development times, leading to a reliance on fossil fuel-powered 2WVs which generate significant greenhouse gas (GHG) emissions and can be a risk factor health issues and mortality. Without a transition to electric 2WVs (2WEVs), LMI families will continue to see increases in these negative externalities as vehicle ownership increases. However, no robust planning tools or methodologies exist to support planning for this transition. In light of this research gap, this paper attempts to identify optimal 2WEV battery swapping station (BSS) sites by first identifying key groups of 2WV riders through clustering demographic and proxy travel diary data collected directly from 2WV riders in Indonesia’s largest urban center, Jakarta. Next, these trip behaviors are translated into an 2WEV context via an agent-based model (ABM) to log hypothetical recharge events. By modeling hypothetical recharge events and subsequently analyzing which urban areas or features they align with using statistical tests, we are able to show which points of interest or demographic factors may be relevant in BSS placement.

ABSTRACT. Global trade is often represented as weighted country-country flow, yet the feasibility of a shipment is conjunctive. Exporter capacity, commodity substitutability, route availability, maritime chokepoints, insurance, and sanctions must all remain feasible. This work models such dependencies as a directed typed hypergraph and asks when the induced cascade law can be reduced to a dyadic network preserving the observed marginals. The obstruction is already present in the smallest nontrivial block. If two alternative supply bundles each have rank r, share q hidden constraints, and each elementary constraint closes independently with probability p, then the hypergraph reliability is R_H = 2(1-p)^r - (1-p)^(2r-q), the dyadic reliability is R_2 = 2(1-p)^r - (1-p)^(2r), and R_2 - R_H = (1-p)^(2r-q) - (1-p)^(2r) > 0. Thus dyadic redundancy is biased upward exactly by shared bottleneck overlap. Geopolitical Cascade Irreducibility is defined as the W_1 distance between final-loss laws on the typed hypergraph and on a dyadic benchmark under the same shock ensemble. To locate the obstruction, Topological Redundancy Loss is defined as the normalized disappearance of persistent H_1 cycle mass in a lifted route-substitution filtration. These cycles are substitutable route loops that persist across detour-cost and risk thresholds. Their loss turns nominal alternatives into common-cause failures. Using public country-level maritime chokepoint-dependency tables, compound stress tests over 132 shock families give Pearson r = 0.81 between topological redundancy loss and cascade-law irreducibility. Hardening one at-risk chokepoint selected by a TRL-GCI score reduces mean final cascade size by 0.629, compared with 0.605 for volume targeting and 0.518 for betweenness targeting. The result is a reducibility test for geopolitical trade cascades and a topological protection rule for preserving true route redundancy.

ABSTRACT. Adaptive threshold models offer a simple way to study how local adoption rules produce large-scale cascades on networks. In standard threshold cascades, exposure below threshold is usually treated as having no lasting effect. This submission studies a small extension in which failed exposure increases future resistance: if a node is exposed to active neighbors but does not activate, its threshold rises. I analyze this model on Erdos-Renyi, Watts-Strogatz, and Barabasi-Albert networks using parameter sweeps over the initial threshold q0 and the resistance strength gamma. Across all three topologies, adaptive resistance reduces the parameter region that supports global cascades. In boundary terms, global cascades require lower initial thresholds, or more susceptible initial conditions, in order to occur. The effect is strongest in the Watts-Strogatz case, where the estimated cascade boundary shifts substantially more than in the Erdos-Renyi or Barabasi-Albert cases. A modest robustness check over initial seed fraction shows the same qualitative pattern. These results suggest that failed exposure should not always be treated as neutral in threshold contagion models, and that simple forms of inhibitory memory can alter cascade transitions in topology-dependent ways.

ABSTRACT. This paper asks whether a single invariant structure underlies human systems from early hominins to contemporary populations. We derive five degrees of freedom (5DoF) required for any system to persist under uncertainty: (1) representation of states, (2) maintenance of viable boundaries, (3) coordination among components, (4) constraint of transitions, and (5) weighting of signals. These degrees of freedom define the minimal control architecture required for persistence and provide a common framework linking physical systems, cognition, and human social organization. We model the human colony as a composite system composed of personbytes, firmbytes, and publicbytes, through which embodied cognition is distributed and coordinated. Using archaeological evidence from one early hominin, Homo erectus, and ethnographic data from one fully nomadic present-day population of hunter gatherers, we show that human systems do not primarily evolve to reduce exposure to environmental threats. Instead, they evolve to buffer viability under persistent exposure through coordinated information processing across agents. This coordination generates stable structures that sustain populations across ecological variability over long time scales. To connect physical constraints to cognition, we integrate Active Inference and Semantic Information Theory. Active Inference formalizes how an agent infers hidden states and selects actions under uncertainty. Semantic Information Theory formalizes which of those inferred distinctions are causally necessary for persistence. Together, they specify a single requirement: persistent systems must both infer states and preserve viability-relevant distinctions. This integration provides an ontologically coherent bridge between inference dynamics and viability constraints. Population-level viability is anchored empirically using replacement fertility rates, which define thresholds below which systems risk collapse. The framework therefore links thermodynamic constraints, cognitive processes, and demographic outcomes within a unified structure. We conclude that the human colony operates as a distributed control system that enables the accumulation of information across generations while maintaining viability under uncertainty. This perspective establishes a foundation for analyzing artificial systems as non-biological instantiations of the same underlying constraints.

ABSTRACT. Extended Abstract It is interesting to note that many scientific thinkers today appear to assume that complexity is a phenomenon pertaining mainly to what are known as the ‘hard’ sciences; so deeply has this notion become engrained in our minds that we have indeed come to take for granted that those who deal with complexity must necessarily be experts in one of the scientific fields included in the above category. In recent years, this concept has been so thoroughly reinforced and introjected that we now find ourselves pondering in what ways complexity can be “applied” to disciplines such as the humanities, the fine and performing arts, literature, music, and so on. And yet, if brought under more accurate perusal, it should become obvious that this approach is tantamount to “doing things backwards”, because it is the arts themselves (of every genre) which are examples of the many emergent properties that arise spontaneously from complex systems, defining, transforming, and (self)-organizing the systems themselves. Even more to the point, it is often these very arts which tend to precede, foreshadow, or even inspire the inventions, scientific discoveries and mentalities that are as yet on the horizon. Examples of this can be found in every culture and every historical period, but for the sake of brevity, taking into account only the fine arts of the last century, it is sufficient to consider the many intuitions into complexity that were expressed by artists such as Picasso, whose multi-dimensional perspectives hinted at the intrinsic relativity of reality, along with impressionists and post-impressionists like Monet or Van Gogh, surrealists like Magritte, or Escher and his strange worlds, representing but a few of the ways in which artists introduced their contemporaries to a different perception of life, at times anticipating scientific breakthroughs that they foresaw or envisaged. Not only the visual arts, but also music, dance, literature, film and so on are all harbingers of the times to come, as well as being the most meaningful icons of conservation of humanity’s past glories and tragedies. When speaking of arts and humanities, we are speaking of the social, and let us not forget, above all, that social (human) systems display exactly the same interconnected, interdependent, interacting, and unpredictable properties as all other complex adaptive systems. Furthermore, while the traditional scientific methods of observation are capable of accurately and objectively observing systems that are merely complicated, of breaking them down into smaller parts and of controlling their evolution, another aspect that must be kept in mind is that it is impossible to observe a complex system from an external standpoint: each observer will affect and be affected by the system; every observer in fact is inevitably an “observer/participant”, as is commonly termed in social studies. One of the differences that stand out in studies on social systems, but which could be confirmed by research on other living complex systems, is the fundamental importance of the qualitative factors, which tend to be neglected in favor of the system’s quantitative factors, easier to obtain and to analyze. It should also be mentioned that the inherent difficulties which are encountered when attempting to describe, explain, formulate, or physically illustrate complexity, or the emergent properties of complex systems, are related to the fact that any kind of graph, explanation, description or figure we attempt to create automatically becomes part of the complexity of the system itself, encompassing therefore all of the uncertainty and non-measurability that are intrinsic to complexity. Here, once again, it is the humanities, literature and the arts which can come to our aid: not only as precursors to our discoveries and evolutions, but also as the best way that we can hope to illustrate the nodes and interconnections of complexity, not only with images, but also with words, keeping in mind that the words and images we choose transform the way our mind perceives itself. Perhaps for this very reason, the visionary projections made accessible through art, music, dance and so on are the most effective forms of knowledge, or symbolic mediation, that can succeed in representing, or at least providing us a glimpse of complexity (or time, for that matter). Likewise, the sole human endeavor which can give us a hint of the scope and characteristics, or even an insight into the future of a complex system, in the sense of concepts that have yet to be formulated or developed, is to be found in the arts, be they visual, performing, auditory or literary. And yet, these have for some time been artificially separated from the sciences, in the name of a dubious (and as Simon would say) limited rationality. Like the attempt to consider technology as separate from culture, the false dichotomy between art and science (which are both expressions of the human capacity for wonder, curiosity, and creativity) has its roots in the same fallacious mindset that values quantity over quality, even in relations and connections; this is one of the hypertechnological civilization’s fatal errors that urgently needs to be remedied.

ABSTRACT. In digital advertising, campaigns operate in dynamic environments where performance metrics such as spend, impressions, clicks, and revenue interact and change over time. However, the evaluation of campaign performance often relies on isolating these metrics or evaluating at specific points in time. Assessing performance like this makes it difficult to determine whether unusual performance reflects meaningful system behavior, inefficiency, or temporary perturbations. This paper presents an early-stage framework for detecting anomalous campaign performance and evaluating whether such anomalies are associated with downstream feedback effects in digital advertising systems. Using two years of daily campaign-level Google Ads data from three anonymized clients, LASSO Regression models were used to estimate expected clicks and revenue within 28-day temporal windows to maintain contextual neighborhoods. Residuals from these models were standardized and combined to create a composite anomaly score. Also, the observations were categorized into four performance categories, indicating the level of engagement relative to revenue attainment. The results from the analysis showed that anomaly patterns varied across clients, indicated anomalous behaviors, and client specific digital advertising systems were not uniform. The feedback analysis suggested stronger relationships between anomaly scores and changes in spend or revenue in shorter-running campaigns, while the relationships weakened as campaign duration increased. The findings suggest that anomaly detection, when combined with time-based feedback analysis, can provide useful insights into campaign behavior, support more contextual investigations, and decision-making in digital advertising.

ABSTRACT. History shows a succession of dominant currencies over long periods ranging from several centuries to several decades. All of them, however, eventually decline. Against a backdrop of multiple uncertainties surrounding the prospects for the dominance of the US dollar, we propose a non-linear graph model for modeling currency dominance and the causes of its regimes switchings. These are based on a multidimensional environment variable that includes shocks related to technology, development, monetary and fiscal institutions, geopolitical forces and conflicts, and the regulatory environment, for which we provide an original database of events. Conditions based on this environment variable determine, via conditional Markov chains, the dynamic structure of the IMS. The model is calibrated for nine reference currencies for the period stretching from the beginning of the Classical Gold Exchange Standard to the present day. Having validated the calibration, we proceed in three steps. First, we develop and code country hypotheses deemed most likely in the nearby future, in the same form than the events database, for the various currencies. Secondly, we combine these hypotheses to form global scenarios. Finally on this basis we simulate the calibrated model for the prospective period, for each scenario. The dominance of the US dollar is affected in some of the scenarios tested, but only marginally, indicating strong resilience in the system’s current centripetal dynamics.

ABSTRACT. Exodromy and Epigenetics: A Galois-SNFT Approach to Dynamics M. Swan1, T. Kido2, R. P. dos Santos3. (1) University College London, Malet Place, London WC1E 6BT UK, melanie@DIYgenomics.org. (2) Teikyo University. (3) academicum.ai. Modeling complex systems requires mapping exponentially branching “many-worlds” trajectories, where each path acts as an encapsulated module assessed for tiny, high-stakes permutations. We develop exodromy (exit paths to higher dimensionality [1]) using a Galois Smartnetwork Field Theory (SNFT) approach to model the systems biology of aging [2]. This framework accommodates high-volume, path-dependent branching, formal verification, and audit logs, specifically addressing how age-related dysregulation in DNA methylation acts as a reversible switch for gene-regulatory network dynamics [3]. The model represents GRN–epigenetic–transcriptomic interactions as a filtration of directed combinatorial complexes, using ramification (branching), monodromy (looping), and exodromy (exit path dependence) (Figure 1). On the discovery side, high dimensional topological holes alert when regulatory pathways diverge (via persistent homology across activation level filtrations). On the assurance side, well formedness is enforced through explicitly path dependent, trajectory specific constructions that guarantee consistency. High Volume Branching via Directed Topological Filtrations: The state space is modeled as a directed simplicial (or directed flag) complex X. Branching structure is quantified by computing persistent homology over a filtration of activation thresholds. Persistence barcodes of the boundary maps ∂_k:C_k (X)→C_(k-1) (X)capture divergence, with directionality preserved through a directed variant of persistent homology. Path Dependency via Directed Persistence Modules: System history is encoded in a directed persistence module M:(R,≤)→Vect, whose structure maps M(s≤t):V_s→V_timpose temporal asymmetry and non invertibility. Exodromy is encoded by ensuring that admissible future states are constrained by the entire past trajectory. Formal Verification and Traceability via Invariants: Each persistent feature αis assigned a stable identifier derived from its start–end pair (s_αⓜ,e_α )and its homology class representative [z_α]∈H_k (X). The tag is invariant under chain homotopy equivalence, enabling an auditable mapping from starting to ending signatures.

1. High-Volume Branching 2a. Exodromic Exit Paths 2b. Path-Dependent Dynamics 3. Formal Verification Maps

Figure 1: Exodromy Model of GRN-epigenetic-transcriptomic interactions. References [1] Barwick, C., Glasman, S. & Haine, P. 2018. Exodromy. arXiv:1807.03281. [2] Swan, M., Kido, T. & dos Santos, R. P. 2026. Galois Smartnetwork Field Theory for Millennium Prize Math Discovery. AAAI Spring Symposium Proceedings. [3] Moqri, M., Ying, K., Poganik, J.R. et al. 2026. Integrative epigenetics and transcriptomics identify aging genes in human blood. Nat Commun 17(725).

ABSTRACT. This research examines the structural evolution of the Egyptian Exchange (EGX) in response to major systemic shocks including the COVID-19 pandemic, the Russia-Ukraine war, and regional geopolitical conflicts through the lens of Network Event Studies. Moving beyond traditional price-based tests of the Efficient Market Hypothesis (EMH), this study introduces Abnormal Network Metrics to quantify structural deviations in market density, modularity, and centrality. Using a Composite Efficiency Index of sectoral return data of the EGX, daily weighted adjacency matrices were constructed based on absolute sector correlation coefficients. For each of the identified systemic shocks, we measured five core network diagnostics: Density, Average Clustering Coefficient, Modularity, Betweenness Centrality, and Louvain Community counts. By calculating "Abnormal" deviations the percentage change of these metrics on the event date (Day 0) and over subsequent 10, 20, and 30-day windows relative to a pre-event baseline, the study captures the immediate systemic reflex and the subsequent structural realignment. The data reveals that while most shocks induce a rapid increase in modularity as the market fragments into independent clusters, the Gaza invasion stands out as a unique outlier, triggering a 17.9% spike in density and a 55% collapse in modularity on Day 0. Using this multi-scale approach, the results reveal that the EGX behaves as an adaptive system consistent with the Adaptive Market Hypothesis (AMH). While the market typically responds to shocks through protective sectoral decoupling (increased modularity), high-intensity geopolitical events can trigger an instantaneous systemic collapse of modularity and a spike in synchronization, creating hazardous structural bottlenecks. The findings demonstrate that while price-based efficiency may appear restored shortly after an event, structural anomalies particularly in Betweenness Centrality persist for longer periods, signaling latent systemic vulnerabilities. Ultimately, the study confirms the long-term resilience of the EGX, as the network consistently reverts to a modular architecture to prevent cascading failures, providing a novel framework for regulators and investors to monitor systemic risk in emerging markets.

ABSTRACT. Artificial Intelligence (AI) is rapidly entering healthcare systems, yet major limitations have been identified, such as lack of explainability, human trust calibration and overreliance, and poor utility and poor integration. Combined, these risks indicate that healthcare AI is not always appropriate or safe. Much of the current research focuses on improving the performance and accuracy of AI systems, while far less attention has been given to identifying situations where AI should not be used due to major risks or interference with clinical decision-making. In this work, we approach this question from a complex systems perspective by examining how discussions about AI in healthcare emerge and develop within large online communities. To explore these discussions, we analyze large-scale Reddit data containing conversations among healthcare professionals, patients, technologists, and members of the public. Because the dataset includes millions of posts and comments, we designed a computational pipeline that reconstructs entire discussion threads and allows them to be studied at scale. The workflow converts compressed Reddit archives into efficient analytical formats, links posts and comments to rebuild complete conversations, and gradually filters discussions using language detection, domain-specific keywords, semantic embeddings, and large language model classification. Each reconstructed conversation is treated as part of a broader socio-technical discourse system. By representing conversations through semantic embeddings, we can examine how discussions group together around shared concerns and themes. Early observations suggest recurring patterns in how participants talk about the limits of AI in clinical settings. These discussions frequently raise issues related to trust, responsibility in decision-making, system reliability, and accountability. By analyzing these patterns across a large and diverse set of conversations, this study aims to better understand the conditions under which AI use in healthcare may introduce systemic risks. Identifying these boundaries is important for viewing AI not simply as a technological tool, but as part of a complex socio-technical system that shapes clinical practice, governance, and policy.

ABSTRACT. Complex systems respond to perturbations in multiple ways: they may resist change, recover after disturbance, reorganize into a new viable regime, or even improve through exposure to stress. Terms such as stability, resistance, resilience, robustness, adaptation, and antifragility are widely used across ecology, infrastructure studies, network science, learning systems, and socio-technical analysis. However, these concepts are often defined and measured through different variables, scales, and assumptions, making comparison across systems difficult. We propose tolerance landscapes as a multiscale framework for representing functional continuity under perturbation. Instead of treating resistance, resilience, adaptation, and antifragility as isolated labels, we represent them as distinct dynamical regions within a shared landscape of possible system responses. Formally, a tolerance landscape can be written as: $$L = (M, g, \tau, \Pi, D, \Xi)$$ where ($$M$$) is the state space of the system, (g) is a geometric or informational metric, (\tau) is a topological structure, (\Pi) is the payoff or functionality whose continuity is at stake, (D) represents system dynamics, and (\Xi) is a family of admissible perturbations. A perturbation (\xi \in \Xi) may be indexed by intensity, temporal scale, and organizational scale. It induces a deformation of the original landscape, potentially affecting dynamics, payoff, accessibility, connectivity, or viable regions. Within this framework, resistance corresponds to low sensitivity of payoff to perturbation intensity; resilience refers to functional recovery after deformation; robustness indicates approximate preservation of landscape structure; and adaptation involves reorganization toward a new viable regime. Antifragility is treated as a distinct regime: it requires perturbation-induced improvement in the structure of viable futures, not merely change or survival. This improvement can be assessed through changes in payoff, accessibility of viable regions, connectivity among such regions, and reduced exposure to fragile or low-functionality regions. The figure on page 3 illustrates this idea by distinguishing the full tolerance landscape from its projections. The upper panel represents the landscape as a geometric-topological space containing regions of resistance, resilience, adaptation, antifragility, and collapse. The lower panels show how observable dose-response curves and temporal trajectories emerge as projections of that broader landscape. In this sense, a flat response corresponds to robustness, a concave response to fragility, and a convex gain under perturbation to antifragility, following Taleb’s convexity intuition. Information geometry and persistent topology provide operational tools for estimating these landscapes. Fisher metrics may capture distinguishability among system states, while Betti numbers and persistence diagrams can summarize the connectivity and persistence of viable regions. Tolerance landscapes therefore offer a candidate framework for comparing how complex systems maintain, recover, reorganize, or improve functionality across perturbation intensity, time scale, and organizational scale.

ABSTRACT. Complex socio‑technical systems face accelerating polycrisis, yet conventional governance models treat disruption as a deviation to be controlled rather than a catalyst for renewal. This paper introduces the Continuity‑Rupture‑Realignment (C‑R‑R) framework, which reconceptualizes rupture as a Generative Pivot – a time‑bounded window where destabilization enables reflexive learning, agency activation, and systemic transformation. C‑R‑R consists of three phases (Continuity, Rupture, Realignment) and is operationalized through the Pivot Readiness Score (PRS), a metric that bridges micro‑level capacities (learning reflexivity, distributed agency, residual continuity, transformation flexibility) with macro‑level systemic resistance (structural rigidity, systemic tension). When PRS > 1, adaptive realignment is possible; when PRS < 1, collapse becomes likely. Illustrative applications to AI governance, pandemic response, supply chains, and energy transitions reveal that successful realignment depends on high adaptive capacity before rupture. C‑R‑R offers a cross‑domain explanatory mechanism, a diagnostic tool for intervention, and a framework for navigating polycrisis beyond resilience.

ABSTRACT. Small open developing economies can be understood as adaptive systems in which external shocks, domestic state variables, and policy instruments interact through feedback, thresholds, and regime shifts. This study examines Bangladesh as an open economy exposed to oil shocks, remittance shocks, and external demand or trade-price shocks. For tractability, the broader complex-systems map is simplified into an estimable policy-response framework: exogenous shocks affect inflation, exchange-rate or external pressure, domestic activity, and fiscal stress; these state variables then shape the policy rate, FX management, and fiscal stance.

The empirical design uses quarterly data for 2000Q1–2024Q4, with monthly extensions for sharper monetary and external-sector identification. A block-exogenous framework treats Bangladesh as too small to affect global external conditions contemporaneously. The baseline can be estimated with either a small SVAR or local projections with regime interactions. Policy feeds back into domestic state variables mainly with lags, improving identification while retaining adaptation.

The framework translates a broader complex-systems view into an estimable open-economy macroe-conometric design. By simplifying the interaction network to exogenous shocks, policy-relevant state variables, and coordinated policy responses, it becomes feasible to study how Bangladesh’s monetary and fiscal authorities react to external stress while preserving the core logic of adaptation, feedback, and regime dependence. Thus, it connects complex-systems reasoning to open-economy macroeconometrics by treating coordinated monetary-fiscal adjustment as an estimable projection of a broader adaptive system. Figure 1 (see the attached abstract in pdf) summarizes the estimable interaction map.

ABSTRACT. Urban transportation networks in megacities exhibit complex structural properties that critically determine their resilience to disruption. This study employs graph-theoretic methods and network science frameworks to systematically evaluate the resilience and structural integrity of the street grid in Mexico City's historic center — among the most densely packed and traffic-saturated urban environments across Latin America. We model the network as a directed graph $\mathcal{G}(N, E)$ with $N=581$ nodes (street intersections) and $E=1,123$ directed edges (street segments), where directionality explicitly captures the asymmetry of traffic flow patterns. We compute centrality metrics as closeness ($C_c$), eigenvector ($C_e$), and betweenness ($C_b$) to identify structurally critical nodes and edges. We also perform complementary analyses to identify the existence of giant components and global efficiency metrics, which provide a comprehensive characterization of street network performance under perturbation. Our findings isolate a limited cluster of junctions and road segments whose elimination triggers deterioration in overall network connectivity. These findings establish a reproducible, data-driven framework for identifying critical infrastructure and assessing resilience in large urban conglomerates, with direct implications for traffic management and urban planning policy.

ABSTRACT. Despite the foundational role of teams as the structural building blocks of organizations, current organizational science lacks a comprehensive construct to capture a team’s overall standing within the organizational hierarchy. While existing frameworks illuminate facets such as team cohesion, efficacy, or performance, they fail to account for a team’s relative capacity for impact and its recognition by the broader system. This paper addresses this gap by proposing Team Influence, defined as a team’s overall position and capacity for impact, operationalized through the relative volume of organizational support and resources it receives. Grounded in Open Systems Theory, Resource Dependence Theory, and the Attention-Based View, this work theorizes Team Influence as a dynamic, compilation-based emergent construct. It is argued that this influence is driven by the diversity of economic and social capabilities, moderated by strategic concentration and the homogeneity of departmental and organizational climates. To evaluate this framework, an agent-based computational modeling is employed, simulating a multilevel organization across twelve monthly periods under varying conditions of workforce stability, leadership quality, and organizational climate. The model explicitly accounts for temporal lags in two directions: the team’s internal absorptive capacity required to integrate new resources, and the organization’s structural inertia in recalibrating its recognition of changed team capabilities. Results demonstrate how team influence trajectories emerge, diverge, and stabilize, offering a formal theoretical bridge between individual-level attributes and macro-level organizational architecture. This research provides critical implications for understanding intra-organizational inequality and the dynamic processes governing multilevel emergence.

ABSTRACT. Color perception is a complex system where physical stimuli intersect with linguistic interpretation. In the World Color Survey (WCS), "White" and "Black" are defined as universal color anchors [1][2]. Metaphors like "settling things in Black and White" (Japanese) or "heibai fenming" (Chinese) reflect this strong, cross-cultural consensus. This study analyzes the transformation of these anchors by comparing WCS data (110 languages) with modern XKCD Color Survey big data (~220,000 respondents) [3].Analysis based on lightness (L*) identified contrasting evolutionary paths. First, the "Black" region (L* < 10) remains a physical constant across eras, maintaining high consensus with minimal vocabulary (only 17 terms in XKCD) (Fig 1). Second, the "White" region (L* > 90) has fragmented into 110 distinct terms in modern society, deconstructing from a universal anchor into a cultural variable (e.g., ivory, eggshell) (Fig 1). Third, "Purple," once unstable in WCS, has emerged as the rank-1 most popular modern name, shifting the center of perception from physical poles to cultural resolution. This suggests that linguistic evolution is moving from bipolarization based on physical constraints toward multi-polar semantic networks. The world where "settling things in Black and White" was straightforward is transforming into one requiring more nuanced consensus-building due to increasing linguistic sophistication.

ABSTRACT. Urban congestion and incidents cause massive economic losses, yet conventional forecasting fails to capture the irregular fluctuations of accident-induced disruptions. Current approaches suffer from a modality gap, processing sensor data and text records in isolation. We propose a unified multimodal spatio-temporal framework applied to the PEMS-BAY dataset and 14,243 accident records. Each time-location pair is represented as a token packet combining quantized velocity values with RoBERTa-based incident embeddings. The architecture utilizes TinyLlama 1.1B with Low-Rank Adaptation (LoRA), fine-tuning only 0.10% of parameters (1,126K) in under 30 minutes. A fusion layer projects the concatenated embeddings into the Large Language Model (LLM) space. Evaluation on 8,699 test samples shows our model achieves the lowest RMSE across all horizons compared to ARIMA and neural baselines. Notably, the model maintains a stable RMSE of 17.33 km/h across 15-, 30-, and 60-minute windows, while baselines degrade. This stability confirms that semantic context enables the LLM to capture superior temporal dependencies. Training converged successfully with a validation loss of 1.9152. Future work will scale the backbone and incorporate real-time streaming for intelligent transportation management.

ABSTRACT. Several approaches have converged towards a set of connected ideas that use information-theoretic and computational complexity principles to model emergence and causality across scales—here named the emergent multi-scale causality (eMSC) framework. Despite its growing popularity [1], several authors have noted that the framework's methodological and metaphysical import remains unclear [2]; this article argues eMSC overlooks an important aspect of the problem it addresses. Despite differences in specificities and intellectual origins, four pillars indicate a general common thread: P1) Emergence is thought in terms of the capacity to construct effective theories from observations of natural systems (Fig.1a). Given a task T—solvable if a generalization G (from H1 to Hi of Hn) robustly obtains with a noise tolerance R—, finding an effective theory means identifying an abstract or "coarse-grained" scale S with variables and dynamics that robustly predict Hi, avoiding the costs (or unsolvability) of tracking lower dynamics with more degrees of freedom. P2) Inspired by [3]-like points, in which emergent entities are conceived as patterns compressed from regularities in datasets, a "more inductive bang for the buck" metaphysics of science is adopted. Facts about inductive success—for a given T, G obtains or not—and facts about computational complexity—some compressions of G are informationally more efficient than others—, have metaphysical import because they cannot be hallucinated (points and arrow, Fig.1b). P3) Still, this results in a multiplicity of ontologies, or the challenge of providing a scarcity measure against an ontological density that threatens relativism about emergent taxonomies. Since indefinitely many successful compressions, each yielding its own ontology, can be obtained by offering (marginally) different efficient ontologies for a given task (local challenge), or by selecting different tasks (global challenge). P4) eMSC characteristically proposes an indispensability criterion: the relevant taxonomies are those solving the target task at the lowest complexity cost—containing no redundancies (the star in Fig.1b). However, indispensability only solves the local challenge. [4] capture this wittily (p.17): "for every question we ask It, Nature has a definite answer; but Nature has no preferred questions”. Here, eMSC can either provide a solution to the global challenge, or develop an account of taxonomical multiplicity that avoids undermining the metaphysical status of effective ontologies. But current advances of eMSC focus on offering refined informational measures of scale relations in isolated systems, something that richly explores indispensability, but keeps overlooking the global challenge.

ABSTRACT. We extend the information-driven framework of self-assembly by introducing a mapping from sequence to assembly behaviour through structural and statistical intermediates. While AlphaFold provides high-accuracy predictions of folded protein structures from sequence, it does not directly capture how these structures participate in collective assembly processes. To bridge this gap, we propose that folded structures define a set of effective interaction constraints that determine the statistical state φ of assembling systems. Formally, we introduce a mapping "sequence"→"structure"→ ϕ("structure")→F[ϕ]→"assembly behaviour", where φ(structure) represents a set of coarse-grained descriptors derived from structural features such as conformations, surface geometry, interaction anisotropy, symmetry, and flexibility. These descriptors define an assembly state space, in which different conformational and interaction patterns correspond to distinct statistical regimes. We then define an assembly frustration landscape F[φ], which quantifies the degree to which local interaction constraints are satisfied or remain unresolved during assembly. In this formulation, frustration arises not only from energetic incompatibility, but from statistical incompatibility between competing local configurations. Structures that admit multiple nearly degenerate interaction modes generate heterogeneous boundary states and extended transition regions, while structures with strongly constrained interaction geometries lead to rapid locking into uniform domains. Within this framework, variations in sequence induce systematic deformations of the frustration landscape through their effect on folded structure. Homologous protein families therefore define trajectories in φ-space, along which assembly behaviour transitions between regimes of high frustration (heterogeneous, fluctuating fronts) and low frustration (cooperative, deterministic growth). This provides a direct link between sequence variation and emergent organization, suggesting that evolutionary processes tune not only structural stability but also the statistical rules governing assembly. This perspective leads to a testable hypothesis: sequence variation modulates assembly outcomes by shifting the position of the system within a low-dimensional frustration landscape defined by structural statistics. By combining AlphaFold structural predictions with experimentally measured assembly statistics, this framework enables the quantitative study of how molecular design influences collective organization across scales.

ABSTRACT. Cognitive functions of the brain, such as attention switching and memory retrieval, are underpinned by network dynamics in which large populations of neurons spontaneously and flexibly transition among multiple metastable states. In the field of complex systems science, chaotic itinerancy (CI) has garnered attention as a fundamental dynamical system mechanism that drives these spontaneous transitions. However, there remains a lack of understanding regarding how the complex network structures inherent in the real brain, such as small-worldness and scale-freeness, constrain the conditions for the emergence of CI and its transition dynamics. To address this, the present study aimed to elucidate the impact of structural characteristics of neural networks on the manifestation and transition patterns of CI. We constructed recurrent neural network (RNN) models with diverse topologies (Regular, Small-world, Scale-free, Modular, and Random). For each network architecture, we comprehensively evaluated the phase-space dynamics by systematically varying the statistical properties (mean and variance) of the connection weights as parameters (Fig.1). Numerical analyses revealed that in networks where local clustering is dominant, such as Regular and Small-world topologies, sustained spontaneous transitions among metastable states—corresponding to CI—were observed across a broad parameter space. Conversely, in Scale-free networks, which are characterized by hub nodes, network activity showed a strong tendency to localize at the hubs. Consequently, the parameter region capable of sustaining itinerancy among metastable states was significantly diminished. This study demonstrates that while scale-freeness tends to restrict the diversity of dynamics, the local clustering inherent in small-world structures facilitates the formation of multiple metastable states within the system, thereby relaxing the parameter conditions required for CI to occur. These findings suggest that the brain's small-world-like topology, which avoids excessive concentration of connections at hub nodes, serves as a favorable structural condition to ensure the transitions among multiple states that are essential for cognitive functions.

ABSTRACT. Problem: Each of the first six plenary speakers at the 2025 Science of Consciousness meeting modeled brain activity using different techniques and tools: oscillations and Bayesian models, frequency fractals, rhythms and cycles, evolution and fitness payoffs, information, and neural networks. Questions arise: What other systems models can be applied? Can and how do these models interact to emerge to more fully simulate whole complex systems? Hypothesis: Troncale’s Systems Processes Theory (SPT) offers a framework to address these questions: (1) A systems science should be based on “systems processes” (SPs), patterns of interactivity that can be modeled and tested. SPs are isomorphic; the “same forms” are found in each of the natural sciences, physical to living. (2) SPs interact to form networks – “systems of systems processes.” (3) An ontology consists of SPs and their linkages, organized into a general system of SPs. [1] Method: The website, Complexity Explained, is reviewed in terms of a few of Troncale’s hypotheses. “Systems concepts” are reframed as SPs and their features, functions, and examples. Results: SPs are identified. Chaotic processes, fractals, and power-law distribution are added to encompass the concepts. Network, hierarchy, and boundary, normally seen as structures, are SPs in SPT. The beginning of an ontology appears. Discussion: The ubiquity and utility of a particular set of SPs–control theory/cybernetics, information theory, network theory, for example–inspires its proponents to see it as central. Complexity science starts with self-organization and emergence, and with interactions, things in relation, and expands out. An ontology as a system of systems processes facilitates this expansion. From a SPT perspective, “complexity” is a feature, not a systems process, that emerges through the SP, systems ontogenesis – the diversification and synergistic integration of new types of more complex systems from subatomic particles onward. How systems interact to emerge as more complex systems requires a full view of systems of systems processes, and from this, a research agenda for an integrated metascience appears.

ABSTRACT. Lexical networks provide a quantitative way to compare written and spoken structure across languages. Building on an orthographic-phonological multiplex framework, we analyze eight Germanic and Romance languages using two aligned layers: written word forms and International Phonetic Alphabet transcriptions. Following the original study, each corpus was derived from Project Gutenberg texts after removing function words, stop words, and symbols, yielding 50,000 words per language. In each layer, two words are connected when their Damerau-Levenshtein distance is at most l, with l=1,2,3. The layers are constructed independently over the same lexical node set; cross-layer structure is then measured through distributional distances and shared- neighbor overlap. As a value-added extension, we use these multiplex descriptors for supervised language-family classification. For each language, Jensen-Shannon distances between orthographic and phonological degree distributions and average Jaccard overlaps are concatenated across the three thresholds, forming a six-dimensional feature vector. A nearest-centroid classifier with leave-one-language-out validation correctly classified 7 of 8 languages. The only error was French, a Romance language assigned to the Germanic Centroid. This agrees with its unusually low overlap between orthographic and phonological neighborhoods: at l=3, French had average Jaccard overlap 0.212, compared with 0.649, 0.631 and 0.664 for Portuguese, Spanish and Italian. A complementary classifier based on normalized Shannon entropy also reached 7 of 8 correct classifications, suggesting that cross-layer similarity and degree-distribution heterogeneity capture distinct predictive signatures. These results motivate a phonology-aware correction pipeline in which phonological neighborhoods generate spoken-confusion candidates and multiplex overlap helps rerank them.

ABSTRACT. Collective motion in self-propelled particle systems provides a minimal setting for studying how macroscopic order emerges from local interactions. Lane formation in counter-propagating flows is a representative self-organization phenomenon observed in pedestrian crowds and active matter. Most conventional models assume explicit segregating mechanisms, such as interspecies repulsion, anti-alignment, or collision avoidance. This raises a basic question: are such mechanisms essential, or can lanes emerge from local alignment and oppositely directed self-propulsion alone? To address this question, we introduce a Vicsek-type model with bidirectional desired orientation (VMBDO). Each particle is assigned one of two opposite desired directions, while interparticle interactions are purely ferromagnetic local alignment and identical for all particles. No interspecies repulsion or anti-alignment is imposed. We performed numerical simulations in a two-dimensional periodic domain while varying the density ρ and the noise amplitude η, and classified the collective states using polar, nematic, and lane order parameters. The model exhibits five typical states: multiple flocks, a single flock, a lane phase, a two-phase coexistence state, and a disordered phase. In the lane phase, polar order remains small whereas nematic order is high, indicating spatially separated counter-propagating streams. The phase diagram shows a lane phase and a critical density that are absent in the standard Vicsek model, with the flock-lane boundary approximately following η proportional to ρ^{1/2}. These results demonstrate that counterflow lane formation can arise without explicit segregating interactions, as a generic nonequilibrium outcome of local alignment, bidirectional drive, and noise

ABSTRACT. The question of how and why our societies transitioned from egalitarianism to institutionalized hierarchy is among the most enduring problems in the study of human social evolution. Gregory Johnson's foundational theory of "scalar stress" attributes this transition to a need for efficient communication at scale. However, prior work has not yet tested this theory on complex networks. This study extends Johnson's approach by embedding a model society in a network, applying the DeGroot model of consensus formation, and including two orthogonal representations of hierarchy: one using network topology and one using influence weights. The outcome variable is the effect of hierarchy on the growth rate of a society's expected convergence time when subjected to scaling pressure -- derived from the spectral gap of the row-stochastic update matrix. Under this formulation, all 9 conditions demonstrate sublinear growth in coordination costs, directly contradicting the scalar stress theory. In addition, hierarchy only becomes computationally optimal under conditions of disparate influence, which are shown to be universally detrimental to scaling efficiency. Disagreements of model and theory hint at a need to reevaluate the primacy of factors underlying the shift to institutional inequality and the rise of hierarchical social structures.

ABSTRACT. Schizophrenia is increasingly recognised as a systems-level disorder emerging from large-scale dysregulation across interacting molecular networks. However, the organisational principles governing the schizophrenia interactome remain insufficiently characterised. In this study, we employed an integrated complex systems framework combining protein–protein interaction (PPI) network analysis, multidimensional topology, network cartography, and unsupervised machine learning to investigate the regulatory architecture of schizophrenia-associated proteins.

A high-confidence schizophrenia interactome comprising 8660 proteins was constructed using STRING (confidence score ≥ 0.7). Multiple centrality measures were integrated to identify global hub proteins, while K-means clustering and Principal Component Analysis (PCA) were used to detect multidimensional topological patterns. Network cartography based on participation coefficient and within-module degree further characterised the functional roles of hub proteins within the broader interactome.

Intersection analysis across global centrality metrics identified 15 core hub proteins, including AKT1, TNF, IL6, STAT3, MTOR, and BDNF. Unsupervised clustering revealed a distinct hub-dominated topological class comprising 43 proteins with consistently elevated centrality signatures. Cartographic analysis demonstrated that these proteins function not only as local module hubs but also as connector hubs coordinating cross-module communication within the network.

Functional enrichment analysis revealed strong convergence toward neuroinflammatory signalling, intracellular regulatory cascades (PI3K-Akt, MAPK, mTOR), synaptic processes, and metabolic regulation. Collectively, the findings support schizophrenia as a multilayered complex system governed by interconnected inflammatory, signalling, synaptic, and metabolic subnetworks rather than isolated molecular pathways.

This study demonstrates the utility of integrating network science and machine learning for identifying emergent regulatory structures in neuropsychiatric disorders and provides a scalable framework for systems-level analysis of complex diseases.

ABSTRACT. A network is a complex system wherein each edge represents the interaction between two nodes; however, in many cases there is uncertainty in the observed data and thus in the network’s edge values, and this uncertainty will affect network measures calculated using the data. Previous work has shown the effects of node/edge removal/addition on different network measures, but only in the context of missing data used to create binary and often directed networks. Here we derive uncertainty formulae for four node-level network measures – specifically in undirected, weighted networks – which quantify confidence bounds for their ground-truth values, validate these formulae using synthetic data, and analyze the sensitivities of these measures to measurement noise.

ABSTRACT. This presentation presents the Oasis of Fraternity, a Participatory Action Research (PAR) framework designed to empirically operationalize complex thinking through situated civic experimentation in urban public spaces. In the context of contemporary “polycrisis” - democratic erosion, social fragmentation, and ecological instability - the research investigates how participatory infrastructures can make complex relational processes observable and actionable in everyday social life. Based on an ongoing case study conducted in Montpellier (France), our approach conceptualizes the Oasis of Fraternity as both a civic initiative and a living laboratory of social complexity. Fraternity is approached not as a fixed normative ideal, but as an emergent and processual relational category enacted through situated interactions, collective practices, and semantic negotiations. Methodologically, the framework combines ethnographic observation, participatory workshops, audiovisual recording, participatory mapping, and co-analysis sessions. Preliminary findings suggest that fraternity is unevenly distributed across urban space and continuously redefined through discourse and practice. The study further introduces the concept of artifact–attractor coupling to describe how participatory devices simultaneously generate, structure, and reveal social interactions across multiple scales. Rather than presenting stabilized empirical conclusions, the paper offers a proof of concept demonstrating how PAR can function as a performative and reflexive interface for investigating social complexity. More broadly, the article contributes to complexity studies by proposing an intervention-based methodological framework capable of producing the conditions under which relational dynamics become empirically accessible.

ABSTRACT. Healthcare mergers and acquisitions are often evaluated through financial and operational outcomes. While these measures are important, they do not always capture how restructuring is experienced by the people most directly connected to the healthcare system. Organizational change can affect both patients and employees, although those effects may appear in different ways and at different points in time. When a healthcare organization undergoes a merger or acquisition, the impact can extend throughout the entire system. Changes may be reflected in direct patient care, employee experience, and organizational culture. A study conducted in 2023 found that patient ratings changed depending on the type of restructuring and the time period being examined. This project expands on previous work by incorporating employee satisfaction data from Glassdoor for the same facilities and by expanding the sample size. By including employee satisfaction data, this study examines healthcare mergers and acquisitions from multiple stakeholder perspectives. It explores whether organizational change is reflected in both patient experience and employee sentiment, and whether those effects occur similarly across stakeholder groups. The study uses a hybrid human and algorithmic workflow to collect and organize publicly available data. Healthcare merger and acquisition events will be identified through Becker’s Hospital Review newsletters and then manually reviewed for accuracy. Python, Selenium, and web scraping methods will be used to collect patient review data from Google and Yelp, and merged with employee satisfaction data provided by Glassdoor. The final dataset will include restructuring type, facility location, restructuring date, patient review ratings before and after the event, and employee satisfaction indicators. This data will be used to statistically analyze the impact of organizational restructuring on people.

ABSTRACT. Hydrothermal Liquefaction (HTL) is a thermochemical conversion process that uses hot compressed water to convert wet biomass into carbon-based fuels without requiring an energy-intensive drying process. Due to its capability to process high-moisture feedstocks, HTL is a promising waste-to-energy technology. Despite material circularity increasing through the waste valorization into HTL-derived fuels, the use of these fuels may still result in the emission of air pollutants, including nitrogen oxides (NOx), sulfur oxides (SOx), particulate matter (PM), and carbon dioxide (CO2), raising concerns regarding environmental, economic and social sustainability. This study develops an integrated sustainability assessment framework for HTL waste-to-energy systems by combining process systems engineering, air quality modeling, and ecosystem service assessment. First, mixed-integer linear programming (MILP) is employed to determine the optimal and cost-effective selection of air pollution control technologies, such as baghouse filters, cyclones, electrostatic precipitators, and selective catalytic reduction systems, to mitigate emissions from HTL-derived fuel utilization. Subsequently, the spatial distribution and transport of emitted pollutants are analyzed using the Community Multiscale Air Quality Model (CMAQ) model. Based on the simulated pollutant concentrations and regional spatial information, including land cover, economic conditions, and environmental characteristics, site suitability assessment is performed. Finally, ecosystem service-based reforestation strategies are incorporated through an additional MILP framework to determine optimal tree-planting locations for mitigating environmental impacts while considering spatial equity and public health exposure. The proposed framework supports sustainable decision-making for HTL deployment by simultaneously considering waste valorization, air quality management, ecosystem restoration, and community health environmental trade-offs associated with local biorefinery development. By incorporating health risk exposure into the present analysis, this work establishes a foundation for evaluating equity trade-offs associated with HTL deployment such that future work can build upon this framework by examining localized socioeconomic impacts across communities.

ABSTRACT. Increasingly, large language models (LLMs) are being deployed as decision-making agents in economic and social systems, advising on resource allocation and mediating institutional interactions. Despite this, a foundational question remains unsolved about their behavior as adaptive agents. Do LLMs respond to rhetorical framing as a strategic signal, or do they filter it as the noise that rational equilibrium models predict it to be? This question combines thought from game theory, complex systems, and the emergent behavioral properties of AI agents trained on human-generated language.

We present an experimental study in which LLMs participate themselves as agents in a number of well-studied bargaining games, including the ultimatum game and the prisoner’s dilemma. They do so using systematic rhetorical variation. We hold material payoffs constant across all conditions while manipulating offer framing across ten rhetorical dimensions, including fairness appeals, authority signals, threat language, loss and gain framing, precedent invocation, in-group solidarity, step-by-step justification, and best alternative signaling. By isolating rhetorical content from informational content, we test whether LLM agents exhibit framing-sensitive behavior similar to human subjects, framing-insensitive behavior consistent with rational equilibrium, or a distinct emergent response profile specific to their training. We compare models cross-sectionally to assess whether susceptibility patterns are general properties of current LLM systems or vary by architecture. We benchmark these results against human baselines collected via controlled online experiment.

This study contributes to the characterization of LLMs as complex adaptive agents whose emergent behavior under strategic interaction cannot be fully predicted from either classical game theory or direct extrapolation from human behavioral data. Findings carry implications for the use of AI in economic contexts and the validity of LLMs as subjects in computational social science.

ABSTRACT. Humans continuously explore and discover new things, thereby expanding and developing culture, including knowledge. Such discoveries build upon prior findings and, unlike in other animals, are thought to accumulate in an open-ended manner. Because knowledge is constructed not only at the individual level but also collectively, the discovery of new knowledge may arise from exploration at the group level. Understanding the “synergy effect” that emerges from collective exploration is therefore important. However, the conditions under which such synergy appears in the exploration process remain unclear. In this study, we represent each individual’s knowledge as a network and integrate these networks to analyze, through a mathematical model, the conditions under which synergy emerges in group exploration. The results reveal that, depending on the structure of the knowledge networks, synergy arises when individuals with an intermediate level of diversity are connected.

ABSTRACT. International migration forms a complex network of relationships between countries that evolves over time. In this work, we analyze the structure and evolution of the global migration system between 1990 and 2024 by modeling the stocks from the UN International Migrant Stock dataset as a directed and weighted graph, where nodes represent countries and edges capture migrant volumes between origin-destination pairs.

Using network science tools, we characterize global topological properties, community structure, and centrality, and evaluate four hypotheses regarding the determinants of migration: geographic proximity, cultural and linguistic affinity, economic asymmetry, and forced crises.

The results reveal a system undergoing progressive globalization: regional communities persist but lose cohesion, migration corridors increasingly span traditional boundaries, and geographic proximity declines in relative importance as a primary determinant. In contrast, linguistic similarity emerges as the most stable and consistent cultural factor throughout the entire period. Forced migration constitutes the most pronounced structural exception: in the presence of political or humanitarian crises, geography regains prominence and drives large-scale displacements toward neighboring countries, regardless of other factors.

Overall, these findings characterize a global migration system that is heterogeneous, hierarchical, and evolving, whose dynamics reflect a combination of long-term structural forces and short-term disruptive shocks.

ABSTRACT. We present a longitudinal network analysis of the Python package dependency ecosystem, revealing systematic structural transformations over nearly a decade of evolution. Software package ecosystems can be modeled as complex networks whose topology reflects development practices and technological trends. To characterize this evolution, we construct directed dependency graphs for two large-scale snapshots (2016 and 2025) and compute structural metrics, centrality measures, and community organization.

Our results show a massive expansion of the ecosystem accompanied by clear topological shifts. The number of packages increases by more than an order of magnitude, while dependencies grow even faster, indicating strong densification. We also observe an increase in average degree, a decrease in modularity, and a marked reconfiguration of central nodes. Whereas earlier stages were dominated by web-development and compatibility libraries, recent snapshots show data science, scientific computing, and testing packages emerging as structural hubs.

These findings indicate that the Python ecosystem evolves toward a more interconnected architecture and follows growth patterns consistent with established laws of complex networks, including densification and hub reorganization. More broadly, this study provides empirical evidence that software dependency ecosystems exhibit systematic structural evolution and demonstrates the value of network science methods for understanding the dynamics, organization, and robustness of large technological systems.

ABSTRACT. The Bang Pakong River sub-basin in Eastern Thailand is experiencing rapid land-use change driven by the Eastern Economic Corridor (EEC), with major consequences for landscape structure and water-related ecosystem services. In this study, the sub-basin is treated as a complex adaptive social-ecological system, combining scenario-based land-use projections with spatial ecosystem service assessment to explore future sustainability and adaptability pathways using a spatially explicit land system model. Land-use change between 2021 and 2036 is simulated under contrasting Business-as-Usual (BAU) and Sustainable Development Goals (SDGs) scenarios, driven by biophysical and socioeconomic factors. Ecosystem service provision—focusing on habitat quality, crop production, water yield and urban flood mitigation—is evaluated for each scenario using modelled land-use outputs and supporting biophysical and socioeconomic data. The BAU scenario shows continued urban expansion and conversion of forest and miscellaneous land to agriculture and built-up areas, reflecting current EEC development trajectories and intensifying land-use conflicts, whereas the SDGs scenario exhibits lower urban growth and a net increase in forest cover while retaining agriculture as the dominant land-use type. These contrasting trajectories result in spatially uneven ecosystem service potentials, with BAU increasing pressure on water regulation and habitat services and SDGs maintaining or enhancing key services in conservation and climate-adaptation priority zones. Overall, the results illustrate how alternative development pathways can either undermine or support the resilience of land–water systems in rapidly transforming river basins. Framing the Bang Pakong sub-basin as a complex adaptive system helps to identify hotspots where land-use decisions strongly affect multiple ecosystem services and to highlight policy levers that can steer the system toward more sustainable and flexible futures. These insights are valuable for spatial planning and water-resources management in the EEC region and contribute to broader efforts to integrate land-use scenarios and ecosystem service assessments within complex

ABSTRACT. Financial markets, in which many agents interact intentionally, algorithmically, or randomly, are quintessential complex systems. Market orders directly affect future prices, and this influence is called the propagator in the propagator model. To investigate the behavior of the propagator, the linear propagator model, in which price changes depend linearly on past orders, has been used as a foundational modeling assumption. In this model, the propagator is required to decay and satisfy the resilience condition to reconcile persistent order flow with the diffusive nature of prices. However, since the linear assumption is not technically correct in the presence of large metaorder splitting because of the square-root law, careful interpretation of the estimated propagator under the linear framework is required. Here, we perform a sensitivity analysis of the linear propagator framework using a nonlinear propagator model with no decay. We mathematically demonstrate that the estimated propagator exhibits “decay” and satisfies the resilience condition as a consequence of the estimation procedure. This decay can be regarded as a statistical artifact, since our model has no decay. This result suggests that careful consideration of the nonlinearity and model misspecification is required when interpreting power-law decay in studies of propagator estimation.

ABSTRACT. Locusts can rapidly increase in population under specific environmental conditions, forming large swarms that cause severe agricultural damage within a short period. Previous studies on locust control have mainly focused on modeling swarm behavior, predicting migration, and improving early detection through satellite or drone-based monitoring. These studies have contributed to understanding swarm dynamics and improving field surveillance. However, relatively little attention has been paid to how limited intervention resources should be allocated when multiple swarms emerge simultaneously. In such situations, the choice of which swarm to target first becomes an important strategic problem. In this study, swarm growth was modeled as a log-linear process, and damage was formulated as a quantity proportional to the square of the swarm radius. Under this framework, simulations were conducted using different target selection policies, and the effectiveness of each policy was compared using cumulative damaged areas as the evaluation metric. The results showed that, on average, the policy that prioritizes the swarm with the largest radius achieved the greatest reduction in damage. The distinctive feature of this study lies in its focus not on the precise reproduction of swarm dynamics, but on the decision making problem of selecting intervention targets under limited resources. While existing research has emphasized biological realism and monitoring accuracy, this study examines how simple target selection policies alone can influence long-term outcomes. Even under a simplified model, the findings demonstrate that differences in target selection policies can produce statistically significant differences in cumulative damage, highlighting the importance of strategic design in locust control.

ABSTRACT. Pattern formation in network-coupled dynamical systems plays an important role in many physical, chemical, and biological processes [1]. Here, we investigate the emergence of spatial patterns in reaction–diffusion dynamics on quasi-ring networks, obtained through weak perturbations of regular ring topologies. Quasi-rings can be interpreted as discretizations of imperfect continuous media with periodic boundary conditions. While regular rings support spatially extended Fourierlike modes, weak perturbations of the coupling structure break the translational symmetry and induce strong eigenvector localization [2]. We consider a two-species reaction–diffusion system on a network, ϕ'i = f(ϕi, ψi) + Dϕ∑jLijϕj , ψ'i = g(ϕi, ψi) + Dψ∑jLijψj , ∀i, where A and K denote the adjacency and degree matrices, respectively, and L = A − K is the graph Laplacian. Stability of the homogeneous state is analyzed through the Master Stability Function framework by decomposing perturbations into Laplacian eigenmodes [3]. Near the instability threshold, weakly nonlinear theory is employed to characterize the nonlinear saturation and coexistence of unstable modes. In particular, near continuous (second-order) transitions, the nonlinear spatial patterns inherit the structure of the critical Laplacian eigenvectors selected by the linear instability. Figure 1 illustrates the coexistence of localized short-wave instabilities and delocalized long-wave oscillatory modes in quasi-ring networks. The localized mode confines activity to a small subset of nodes, while the extended oscillatory mode promotes global synchronization. Their nonlinear interaction generates exotic dynamical regimes, including localized oscillations, oscillation death, traveling waves, and coexistence states combining stationary and oscillatory regions. These results demonstrate how weak structural perturbations can strongly reshape the spectral structure of nearly regular systems and produce complex localized dynamical patterns [4].

ABSTRACT. We extend the Sakoda-Schelling segregation model by replacing traditional, rule-based agents with Large Language Model (LLM) agents that make residential decisions using natural language reasoning grounded in social context. To our knowledge this is the first application that substitutes the mechanical agents of the Sakoda-Schelling model with LLM-driven agents. We compare LLM agent behavior across four social and two other contexts: Income (High vs. Low), Ethnic (Asian vs. Hispanic), Racial (White vs. Black), Political (Liberal vs. Conservative), two pairs of colors to serve as control groups, and a mechanical baseline (the standard Sakoda-Schelling model). We measure the resulting segregation via a Dissimilarity Index, which measures how dissimilar neighborhoods are in an overall geographical area. We find that mixture-of-experts LLM architectures partially reproduce empirical segregation patterns collected at the census block level and measured across counties, while other architectures do not:

Mixture-of-experts LLM: income seg. < racial seg. < political seg. Empirical: income seg. < political seg. < racial seg.

ABSTRACT. The mechanism for complex systems to have nontrivial macroscopic states is an active topic in statistical physics. The system of identical Kuramoto-Sakaguchi oscillators with phase shift $\alpha$ is one of the non-equilibrium models that exhibit various patterns. While this model has been investigated in lattices and regular graphs, the key factor governing nontrivial states beyond synchronization in the network geometry remains relatively unexplored. Here, we obtain the phase diagram of the identical Kuramoto-Sakaguchi model regarding the spectral dimension $d_s$ and $\alpha$ on the complex networks. Our numerical simulations demonstrate that the critical value of $\alpha$ ($\alpha_c$), above which (below which) the steady state is a nontrivial one (synchronization), monotonically depends on $d_s$. This phase transition emerges even in finite Gaussian white noise: We theoretically describe the presence of $\alpha_c$ using the renormalization group analysis for a representative graph. Our study thus sheds light on the role of $d_s$ in nonlinear and non-equilibrium systems beyond equilibrium.

ABSTRACT. Water markets under scarcity conditions are inherently complex systems, where the interaction of heterogeneous agents, uncertainty, and strategic behavior gives rise to nonlinear dynamics and emergent phenomena. However, their analysis remains largely dominated by traditional econometric models which, due to their deterministic structure, are inherently limited in their ability to capture such emergent dynamics and adaptive decision-making processes. This limitation calls for alternative methodological approaches capable of representing the true complexity of these systems. In this context, this paper proposes a methodological advancement in Agent-Based Modeling (ABM) by introducing a probabilistic calibration and validation framework specifically designed for complex socio-hydrological systems. The model departs from deterministic behavioral rules by representing agents’ willingness to buy and sell as stochastic, elasticity-based functions, grounded in isoelastic (Cobb–Douglas-type) formulations. This approach enables an explicit representation of bounded rationality, uncertainty, and heterogeneous responses to economic and environmental signals. Agents, including farmers, mining companies, and investors, interact within a decentralized market where transactions emerge endogenously from probabilistic decision processes. Methodologically, the paper integrates three key components: (i) micro-founded probabilistic behavior, (ii) calibration through Genetic Algorithms to efficiently explore high-dimensional and non-convex parameter spaces, and (iii) validation based on the reproduction of stylized facts and statistical distributions, rather than point-wise trajectory fitting. Applied to the Copiapó groundwater rights market in Chile, the model successfully reproduces observed features such as price dispersion, low transaction frequency, and speculative dynamics. Beyond the case study, the main contribution lies in positioning ABM as a rigorous experimental platform for the probabilistic analysis of complex systems, advancing methodological standards for calibration and validation and enabling more robust policy exploration under uncertainty.

ABSTRACT. Peru's 2024 National Assessment of Learning Achievement (ENLA) pairs sixth-grade academic scores with a socioemotional self-regulation module and an extensive student questionnaire on motivation, classroom climate, wellbeing, and family context. Using a framework that allows for latent confounding, we ask how academic language, academic mathematics, and socioemotional learning are jointly structured. Survey-weighted Pearson correlations provide descriptive measures of association, and constraint-based causal discovery with Really Fast Causal Inference (RFCI) estimates a partial ancestral graph that encodes both directed relations and uncertainty from unmeasured causes. Analyses incorporate the Ministry of Education's latent constructs, calibrated sampling weights, and five plausible-value draws for socioemotional learning. A PC-stable baseline assuming causal sufficiency yields a dense graph; relaxing this assumption with RFCI (Figure 1) reveals that academic language and mathematics form a tightly coupled dyad embedded in a broader socioemotional neighborhood, while subjective wellbeing emerges as a robust directed parent of socioemotional learning across all robustness checks. Motivation, perceived classroom climate, violence exposure, and family relational constructs repeatedly act as shared influences adjacent to each outcome. The findings highlight that academic and socioemotional outcomes are embedded in overlapping motivational and contextual structures, suggesting that policies targeting wellbeing, classroom participation, and safety may affect multiple outcomes simultaneously.

ABSTRACT. In many real systems a visible layer of collective oscillations coexists with a hidden interaction layer whose topology is not directly observable. How the hidden topology shapes visible-layer dynamics depends on the nature of inter-layer coupling, yet the diffusive coupling assumed in most multiplex studies is frequency-independent and lacks resonance-based selectivity. Here we couple the two layers inertially, through an off-diagonal mass matrix, so that interaction occurs via accelerations and is frequency-dependent. Formal elimination of the hidden linear oscillator layer produces an oscillatory memory kernel whose eigenmodes and frequencies are governed by the hidden-layer Laplacian. Because higher-degree nodes in the hidden layer project more strongly onto high-eigenvalue Laplacian modes, their kernel frequencies shift away from the visible-layer resonance, weakening resonant energy exchange. The hidden topology is thus imprinted on the visible layer as a non-Markovian oscillatory memory that modulates coherence node-selectively. Numerical experiments on Erdos-Renyi hidden layers (N = 150) confirm that this selectivity emerges only above a threshold in inertial coupling strength, consistent with the avoided-crossing spectral gap exceeding mode damping, and requires nonzero hidden-layer network coupling. Under diffusive inter-layer coupling the kernel loses its oscillatory character, so the selectivity vanishes. These results identify the spectral structure of a hidden-layer Laplacian as a mechanism for node-selective resonance in multiplex oscillator networks.

ABSTRACT. How does green industrial transition reconfigure the urban and structural organisation of the economy under persistent disequilibrium, and can aggregate gains coexist with territorial divergence and selective exclusion? We address this through a recursive micro-meso-macro framework where cities (MESO level) are reinstated as selection environments that filter MICRO level interactions into observable signals (wage ladders, specialisation indices, prices) conditioning bounded-rational planning and interacting with other cities at MACRO level. The economy is a sequence of system states linked by layer-specific transition operators for each level [1, 2]. The framework is implemented through an agent-based model of green structural transition across 30 cities with identical initial conditions, populated overall by 10,000 households and 2,000 firms across four sectors (Green/Non-green × Prototype/Mature). Identical initial conditions provide a clean identification strategy: any spatial differentiation that emerges from micro-level interactions is attributable to the recursive architecture itself. Four results stand out (Fig.1). First, the model generates a two-phase price trajectory in which a sustained green prototype scarcity premium emerges from MICRO-level directed demand reorientation interacting with localised labour-market tightening at the MESO level. In addition, the MACRO level inter-city trade topology locks in early and unconditionally around green production while labour mobility and firm relocation remain structurally volatile, a regime of selective rigidity in which goods flows crystallise before the wider territorial hierarchy has settled in the MICRO and MESO levels. At the MACRO level, the transition produces a dual-inequality paradox: system-wide household income inequality falls while the inter-city component rises, and the intra-city inequality component decreases at the MESO level. The system also exhibits layered convergence: production dominance, network architecture, the territorial hierarchy, and the price structure each reaches quasi-stationarity at different horizons, a property that emerges from the interaction of level-specific operators rather than from imposed exogenous convergence speeds. These results show that aggregate gains, territorial divergence, and network reorganisation are coupled outcomes of a single recursive multi-level architecture stabilizing at different horizons. Policies should consider these different temporalities.

ABSTRACT. Governing equations are fundamental for describing and predicting dynamic urban geographic systems. Unlike physical systems guided by first principles, urban spatiotemporal phenomena emerge from coupled geographic processes that lack deterministic theoretical foundations, making the discovery of governing equations elusive and largely heuristic. Spatiotemporal dynamics in urban systems are often observed as sequential snapshots of spatial distribution, while the cause of such dynamics is often implied or unknown. In this study, we propose a unified differential equation formalism that decomposes urban dynamics into two components: a spatial evolution term, which captures time-invariant spatial interaction processes among locations, and a temporal dynamics term, which represents self-dynamic or periodic changes within locations. Building on this formalism, we introduce the Urban Discovery Framework (U-Discovery), which integrates Hypothesis Generation, Neural Fitting Examination, and Governing Equation Identification for the discovery of governing spatial interaction laws. U-Discovery leverages Large Language Models (LLMs) and literature-based reasoning to propose differential equation candidates. Each candidate was calibrated from the observed spatiotemporal dynamics using a proposed neural fitting method, Urban Differential Equation Network (UrbanDE-Net). The candidates are evaluated and ranked based on the fitting error and mathematical complexity. Our synthetic experiments prove that U-Discovery can find the sole governing equation from the simulated dynamics. Empirical experiments using human activity dynamics in Hennepin County, Minnesota, further show that U-Discovery can identify interpretable governing laws from real-world observations. A newly discovered gravity-model variant, which incorporates edge betweenness, outperforms classic spatial interaction baselines, suggesting that urban activity redistribution is shaped by both distance decay and structural effects of the geospatial network. Additional scale experiments show that candidate rankings shift across spatial resolutions, highlighting the scale dependence of spatial interaction processes.

ABSTRACT. Cities are complex adaptive systems whose resilience depends on social infrastructure — the parks, community spaces, places of worship, and locally oriented businesses where social ties form, information diffuses, and residents take action. Empirical work increasingly links these anchor sites to neighborhood-scale outcomes from disaster recovery to civic participation, yet data-driven measurement has lagged behind the complex geospatial systems required. Methods remain embedded in one-off scripts, outputs are static maps difficult to inspect or compare, and the field lacks the open platforms common in analogous complex systems domains.

To fill this gap, we present socialinfrascorer, an open web-based dashboard that operationalizes the Social Infrastructure Scorecard framework for neighborhood-scale measurement. This platform makes scoring and comparison accessible to planners, researchers, and community members without GIS expertise - at scale, for any urban system worldwide. The dashboard scores neighborhoods across 3 metrics (density, diversity, and spatial dispersion) combined into a composite A–E grade, and integrates spatial data storage, scoring pipelines, a documented REST API, and a map-first interface into a single platform. We document the architecture and data flow, describe distinct user workflows, and discuss the constraints of building a decision-support system on complex urban data. By lowering the barrier to consistent, comparable assessment and enabling crowdsourced expansion of the spatial inventory, our platform supports broader empirical study of how social infrastructure shapes the collective behavior and adaptive resilience of urban systems.

ABSTRACT. Life is thought to have originated on Earth after the appearance of self-replicating molecular assemblies in the primordial soup. Self-replicators, as a product of natural selection, gave rise to agents that exert causal power in the world, but what is the causal structure of a medium before self-replicators appear, and evolution takes hold? Recent advances in causal information theory have advanced our ability to quantify when a system is a causal agent. Indeed, previous work has suggested that what bootstraps life is the ability of simple chemical networks to raise their Φ, which measures the degree the "whole is greater than the sum of the parts". Here, we used the same tools to study the causal properties of an active medium during the appearance of self-replicators. We chose the popular GARD model, often used to study the origin of life on Earth, which simulates assemblies that undergo growth-fission generations by the accretion of environmental molecules. Our analysis across 100 independent replicates found that Φ significantly predicted the initial appearance of self-replicators. Then, we implemented an intervention policy that added/deleted molecules to maximize Φ while preserving the unperturbed GARD dynamics. This experiment significantly increased self-replication persistence (Mann-Whitney, p<0.001) relative to the original runs, suggesting that Φ may be a functional control knob to steer the degree of self-replication of an active medium, before evolution begins to operate.

ABSTRACT. Susceptible components of the Earth’s climate—known as tipping elements—are characterized by critical thresholds, the crossing of which after periods of sustained change can cause sudden perturbations within the local system that lead to destructive effects on the global climate and cascade to other tipping elements. Among these are the Arctic sea ice, which has experienced substantial reductions in recent decades due to accelerating albedo feedbacks and circulation changes, and the Atlantic meridional overturning circulation, which has slowed in recent decades and is at risk of collapse. To qualitatively assess recent changes in the oceanic dynamics of these regions and identify likely modulators, we draw on the framework of complex networks, which has a young yet rich history of applications in climate science. In our approach, nodes represent geographical locations and edges represent causal relationships with respect to sea surface temperature anomalies. We evaluate changes in oceanic connectivity across the Arctic and Atlantic regions using directed metrics and simulations to better understand the effects of global warming on circulation patterns, emphasizing causality by implementing transfer entropy. Local metrics, such as regional degree centrality, betweenness centrality, and final flows, provide regional information on connectivity and influence, while global metrics, such as cycle counts and group closeness centrality, serve as aggregate measures of internal feedback and collective forcing. Preliminary evidence suggests that the directed topologies of both the Arctic and Atlantic circulations have exhibited notable trends and short-term variability, as well as large-scale shifts in external modes of influence. Our findings underscore the pivotal role that such changes can play in triggering tipping points, with profound implications for Earth's climate stability.

ABSTRACT. Chronic diseases are rarely isolated incidents. They emerge from social determinants and exhibit patterns shaped by behavioral, clinical, and structural interactions across racial groups over time. In New York City, historically marginalized populations have experienced disproportionately high chronic disease burden. Understanding these interactions is imperative to developing policy to address health inequities. This study applies a complex systems and computational epidemiology framework to examine disease co-occurrence among Black and Hispanic New Yorkers from 2002 to 2020. It uses the New York City Department of Health’s Community Health Survey (CHS), which began in 2002. A harmonized dataset spanning 19 years was constructed, yielding 155,256 observations across Black, Hispanic, White, Asian, and Other racial populations. Variables included primary care access, insurance status, physical activity, alcohol use, binge drinking, smoking status, hypertension, diabetes, high cholesterol, and obesity. Weighted prevalence estimates were calculated by race and year. Race-stratified comorbidity networks were constructed for chronic disease and behavioral variables using pairwise binary associations. NetworkX and computational graph layouts were used to develop visualizations examining disease connectivity and network structure across racial and ethnic populations. Black respondents displayed the highest prevalence of obesity, hypertension, and diabetes, despite having a relatively high percentage of primary care access and insurance. This suggests that historical barriers persist beyond the availability of healthcare alone. Hispanic respondents displayed high obesity and diabetes prevalence with low primary care access, indicating potential barriers to preventive and longitudinal healthcare access. These findings support the view of chronic disease burden as an emergent phenomenon shaped by behavioral, clinical, and social processes that interact over time.

ABSTRACT. How much work can an agent extract from a temporal sequence of quantum states when it can only operate online under causal constraints - deciding which energy extraction method to use with knowledge of what it has observed before? Here, we study this problem in the context of quantum state sequences that are potentially non-Markovian - generated by some underlying hidden Markov machine that the agent cannot observe. By leveraging techniques from dynamic programming and computational mechanics, we present a method for identifying the provably optimal agent strategy, with time complexity that scales linearly with sequence length. This motivates us to introduce the maximal work such agents extract - \emph{time-ordered free energy} (TOFE) - as a fundamental measure of free energy available in a temporally correlated quantum system subject to causal considerations.

ABSTRACT. Multistatic microwave imaging (MWI) is ill-posed, and its resolution depends strongly on the sensing geometry [1]. We ask whether removing antennas can improve imaging—i.e., whether the system exhibits antifragility rather than mere fragility or robustness [2]—and interpret antenna dropout through a network lens, with antennas as nodes and Tx–Rx measurements as edges, so dropout corresponds to node deletions in the measurement graph [3]. For a closed ring of Nₚ = 16 probes at f ∈ {0.75, 1} GHz, we assemble the multi-frequency imaging operator under the Born approximation and factorize it via energy-truncated SVD (η = 0.99), obtaining the explicit point spread function PSF(·; θ, β) = P_β e_θ with projector P_β = V_{t,β} V*_{t,β}. Resolving capability is Q(β) = 1/FWHM(β), with FWHM the radial −3 dB radius. We report three relative functionals—single-point (SP) at the DOI centre, mean-of-ratios (MF), and average-then-ratio (AG) over a small target set—and label ΔF(β) = F(β) − 1 > 0 as antifragility. All 120 two-antenna removals are evaluated, with delta diagnostics on truncation depth ΔN_t and worst-case sidelobe change ΔSLL_max to rule out artifact-driven gains [4].

ABSTRACT. Our social environments influence nearly every aspect of our lives. Beyond just the quality, instability in the social environment has been linked to greater risk of psychopathology and aberrant brain connectivity. To assess the intersecting associations between social environment change, mental health, and brain network organization, we cluster thousands of subjects based on 10 measures of the social environment recorded over four years. The resulting clusters represent dynamic social envirotypes, or distinct patterns of change in the social environment. We then demonstrate that these dynamic social envirotypes are associated with both distinct mental health outcomes as well as distinct changes in mental health over adolescence. Dynamic social envirotypes also exhibit dissociable patterns of brain network connectivity. Moreover, we find that the patterns of connectivity associated with better mental health look different for different social environment dynamics, suggesting there is no one pattern of brain connectivity that is ‘healthy.' This work emphasizes the importance of brain-environment coupling when it comes to mental health.

ABSTRACT. Tenure is a cornerstone of the US academic system, yet its relationship to faculty research trajectories remains poorly understood. Conceptually, tenure systems may act as a selection mechanism, screening in high-output researchers; a dynamic incentive mechanism, encouraging high output prior to tenure but low output after tenure; and a creative search mechanism, encouraging tenured individuals to undertake high-risk work. Here, we integrate data from seven different sources to trace US tenure-line faculty and their research outputs at a remarkable scale and scope, covering over 12,000 researchers across 15 disciplines. Our analysis reveals that faculty publication rates typically increase sharply during the tenure track and peak just before obtaining tenure. Post-tenure trends, however, vary across disciplines: In lab-based fields, such as biology and chemistry, research output typically remains high post-tenure, whereas in non-lab-based fields, such as mathematics and sociology, research output typically declines substantially post-tenure (Figure 1). Turning to creative search, faculty increasingly produce novel, high-risk research after securing tenure. However, this shift toward novelty and risk-taking comes with a decline in impact, with post-tenure research yielding fewer highly cited papers. Comparing outcomes across common career ages but different tenure years or comparing research trajectories in tenure-based and non-tenure-based research settings underscores that breaks in the research trajectories are sharply tied to the individual’s tenure year. Overall, these findings provide an empirical basis for understanding the tenure system, individual research trajectories, and the shape of scientific output.

ABSTRACT. Mobility, urban structure, and prosperity are deeply connected. City accessibility depends on both amenity locations and the street networks connecting them, yet most accessibility measures focus solely on spatial density or proximity, overlooking network efficiency. Amenities can be close ``as the crow flies" but far to travel to via the street network. To address this, we introduce Pairwise Circuity ($PC_{\alpha\beta}$), a novel measure quantifying the average excess travel distance imposed by the street network when routing between different amenity classes. Specifically, the measure compares the shortest street network path, $d_G$, to the direct Euclidean distance, $d_E$, from an amenity of class $\alpha$ to the nearest amenity of class $\beta$: \begin{equation*} PC_{\alpha\beta} = \frac{1}{n_{\alpha}} \sum_{i \in \alpha} \left[ d_G(i, j^*(i, \beta)) - d_E(i, j^*(i, \beta)) \right] \end{equation*} where $n_{\alpha}$ is the count of amenities in class $\alpha$, and $j^*$ is the spatially closest amenity of class $\beta$. Applying this framework to 371 cities worldwide using OpenStreetMap data, we generate comprehensive circuity profiles for each city. Our analysis reveals that lower circuity indicating more efficient routing strongly correlates with urban prosperity, including higher GDP per capita, better Quality of Life scores, and lower income inequality. These correlations hold robustly against spatial null models, confirming that PC captures topological realities beyond simple density. Furthermore, PC distributions exhibit clear separations between ``Matured" (Global North) and ``Developing" (Global South) cities, establishing network efficiency as a key structural signature of urban development. Ultimately, the PC framework provides a robust, interpretable tool for planners to target network inefficiencies and improve equitable urban accessibility.

ABSTRACT. We present a progressive research line identifying criticality signatures in X (Twitter) data collected during mass mobilisation events. Analysing hashtag co-occurrence networks across multiple demonstrations, we find a robust structural transition at peak activity: a shift from modular organisation toward a hierarchical nested structure. An entropic analysis of hashtag frequency distributions shows that this transition can be localised in time without network construction. Applying Degree-Thresholding Renormalisation (DTR), we recover scale-invariant, self-similar organisation exclusively at the Critical Temporal Window, in direct analogy with renormalisation group approaches in statistical physics. Finally, retweet networks, encoding direct interactions between users, independently recover the same criticality signatures, grounding the phenomenology in an interacting particle system framework.

ABSTRACT. The arboreal turtle ant Cephalotes giniodontus has been shown to optimize its foraging trail networks by minimizing path length to food sources and eliminating cycles [1]. Toward understanding what biological mechanism underlies this optimization, we propose and study a compartmental model for the foraging dynamics of C. goniodontus that gives rise to oscillations in the ant flow. First, we arrive at a simplified model taking onto account only direct interactions between ants that may travel on trails to search for food and ants that remain at the nest. We show that in this model, all trajectories are attracted to a unique equilibrium. We then add mixed-feedback pheromone dynamics to the model, showing that it now admits sustained oscillations in ant flow. These dynamics feature positive feedback from pheromone deposition, which increases the return rate of ants searching for food, resulting in more pheromone deposition, and negative feedback from pheromone evaporation. This analysis provides insight to a possible mechanism used by turtle ants to optimize their trail networks.

ABSTRACT. Institutional signals, such as the passage of policy or high-profile court rulings not only change the law, they also signal social norms and values about specific issues. Recent empirical studies have shown that such signals can rapidly alter perceptions of social norms and behaviors. However, whether perceived norms change depends critically on the perceived legitimacy of the institution, and accompanying public responses. While prior belief dynamics literature has focused on the role of peer interaction in opinion formation and norm perception, the role of institutions in these dynamics remains less understood. In this mixed-methods study, we integrate a mechanistic agent-based model with empirical survey research to investigate how institutional signals, social norms, and institutional trust co-evolve. In our model, individuals update two coupled quantities after observing a policy event: (i) their belief about the social norm, inferred from the institutional signal and public discourse, and (ii) their trust in the institution, which depends on the (mis)alignment between the social norms they perceive and the institutional action. We parameterize the model using empirical survey data, allowing heterogeneity in individual characteristics across the agent population. Our modeling framework allows us to systematically analyze feedback processes between institutional trust, perceived social norms, and political expression, which are difficult to isolate using empirical observations alone. Combining agent-based modeling with empirical survey data, we identify conditions under which policy events catalyze collective shifts toward pro-social norms, and conditions under which they instead erode trust in democratic institutions.

ABSTRACT. To predict and prevent critical transitions in networked systems, effort has been devoted to detecting early warning signals (EWSs) that precede such transitions. Most previous studies have focused on EWSs in additively coupled networks, whereas critical transitions arising in diffusively coupled systems, such as synchronization, are also widespread in real-world settings. In addition, EWSs under partial observation, that is, when only a subset of nodes in a network is observable, have been studied but not been established. Motivated by these gaps, we investigate performance of EWSs extracted from partial observations of diffusively coupled oscillator systems in anticipating impending synchronization transitions using the stochastic Kuramoto model on networks. Our main findings are twofold. First, the rise of the variance of global and local order parameters provides a useful EWS for synchronization if the network structure and natural frequencies are known and the cluster of nodes that synchronizes first is properly identified. Second, even without such information, phase time series of all oscillators at a single weak coupling strength can be used for identifying a subset of nodes that provides EWSs for early synchronization.

ABSTRACT. Contact matrices are core components of compartmental epidemic models. Survey-based contact matrices remain the gold standard[1], but are expensive and, consequently, can be repeated infrequently and often geographically coarse. The proliferation of large-scale mobility datasets [2] has prompted a natural question: can mobility signals substitute for, or augment, survey-based contact matrices? We explore this question by systematically comparing contact matrices derived from mobility data against their survey-based counterparts across multiple socio-demographic stratifications. Our preliminary results suggest a dimension-dependent pattern: mobility-based reconstructions appear to track diary-based references more closely for some socio-demographic dimensions, such as race and ethnicity (Figure~1), than for others, such as age. We discuss how this pattern may relate to the proportional-mixing assumption underlying mobility-based constructions, under which assortative structure can emerge only to the extent that local spatial units are demographically heterogeneous along the dimension of interest. In this contribution, we compare alternative approaches to generating socio-demographically stratified contact matrices from physical mobility data and discuss the conditions under which mobility signals may complement survey-based mixing patterns.

ABSTRACT. Financial markets are complex adaptive systems whose return distributions, correlations, and sectoral interactions change under stress. Standard mean-variance portfolio optimization relies on stationary historical estimates and is therefore vulnerable to estimation error and structural breaks. This study evaluates a Wasserstein distributionally robust optimization framework for portfolio allocation under nonstationary market conditions. By constructing an ambiguity set around the empirical return distribution using optimal transport, the proposed model optimizes against plausible distributional perturbations rather than a single historical estimate. The resulting formulation introduces an L2-type regularization effect that discourages unstable concentrated positions and reduces turnover relative to classical robust optimization. Using a 2015–2025 backtest with transaction costs, the Wasserstein robust portfolio is compared with equal-weight, nominal mean-variance, and elliptical robust baselines. Results show that the Wasserstein model produces smoother sector rotation, lower sensitivity to mean-estimation noise, and improved resilience during the COVID-19 structural break. These findings suggest that distributional robustness provides a useful framework for portfolio construction in adaptive, nonstationary financial systems.

ABSTRACT. The most remarkable aspect about the immune system is that it has no brain, no central controller, and no pre-loaded manual. And yet it learns. It distinguishes friend from foe across billions of cellular encounters, remembers threats it faced decades ago, and occasionally, tragically, turns against the very body it protects. Explaining how any of these works at the molecular level has occupied immunologists for over a century. I argue that a complementary explanation is long overdue. One that starts not with molecules, but with the network they form. We model the immune system as a weighted, directed graph G = (V, E), where nodes represent immune cell populations and molecular mediators, and edge weights encode the strength of interactions updated dynamically by antigen exposure. This is not merely a metaphor. Lymphocyte co-stimulation networks empirically exhibit scale-free degree distributions and small-world connectivity [1], the same structural fingerprints found in power grids and social networks, associated with striking resilience to random failure but sharp vulnerability to targeted attack. Clonal selection, in this framing, is preferential attachment. Clones that bind antigen well accumulate connections, pulling the network toward configurations that clear the pathogen faster. Affinity maturation works much like gradient descent. Through iterative rounds of somatic hypermutation and selection, the immune system progressively refines the fit between receptor and antigen, arriving at a level of specificity that no single cell could have reached on its own The standard account of immunological memory emphasizes long-lived plasma cells secreting antibodies for years, memory B and T cells waiting to be recalled. We do not dispute this, but we think it is incomplete. When antigen clears, the network does not simply reset. Certain edges remain strengthened, connection patterns shift durably, and the graph that re-emerges is structurally different from the one that existed before infection. This is memory encoded not in any single cell, but in the topology of the network itself, analogous to long-term potentiation in neural circuits [2]. The prediction that follows is concrete and testable. Memory loss, in conditions like aging or HIV, should correlate with measurable network rewiring detectable through single-cell transcriptomic reconstruction (see Figure 1). The same framework makes immune pathology interpretable in a new way. Autoimmunity is not simply the immune system "attacking itself". It is a network that has lost critical inhibitory connections, allowing self-reactive clones to activate without restraint. Immunodeficiency reflects a graph too sparse to reliably amplify a coordinated response. Chronic inflammation is a system stuck. Perturbed out of its resting state, unable to find its way back to homeostasis, dynamically, a limit cycle rather than a stable equilibrium [3]. Each failure mode points to a different kind of intervention: not blocking a single cytokine or receptor but reshaping the connectivity patterns that allow dysfunction to persist. We see three concrete directions from here. First, network-targeted therapeutics. Drugs or cell therapies designed to reinforce regulatory hubs rather than silence individual pathways. Second, patient-specific immune network models built from high-dimensional single-cell data, enabling prediction of individual inflammatory trajectories before symptoms emerge. Third, and perhaps most unexpectedly, conceptual transfer: the immune system has been solving decentralized learning problems, robustly, adaptively, without supervision, for hundreds of millions of years. Understanding how it does so may have as much to teach machine learning as machine learning has to teach us about immunity.

ABSTRACT. Global food insecurity is often discussed as a problem of scarcity, but many breakdowns in nutrition access are better understood as failures of connection, timing, and concentration. This study uses the global wheat trade system as a complex network to ask a practical question: how can we identify where a disruption is most likely to become a health-equity problem, and how can that knowledge guide local interventions before the next shock arrives?

I model countries as nodes in a wheat trade network and trade relationships as weighted links. Vulnerability is measured through two simple but interpretable dimensions: net import dependence and supplier concentration. Countries are most vulnerable when they rely heavily on imported wheat and obtain that wheat from a narrow set of suppliers. This framework produces a two-axis vulnerability map, where the upper-right quadrant identifies countries exposed to both high dependence and low diversification. I then examine how countries move through this vulnerability space over time, treating vulnerability not as a fixed condition but as an evolving balance between robustness and adaptability.

The 2022 Russia-Ukraine wheat shock shows why this matters. Russia and Ukraine accounted for about one-third of global wheat exports, and the invasion triggered wheat price spikes and food-security concerns in import-dependent countries. FAO also warned that concentrated export supplies, high fertilizer and fuel costs, port closures, and logistical disruptions could intensify global food and nutrition risks.

The contribution of this project is not only descriptive. By identifying countries and communities exposed to trade fragility, complexity science can help design smaller, targeted buffers: diversified sourcing, food recovery partnerships, waste reduction, and low-stigma distribution models such as Daybright-style food lockers. These local interventions will not solve the global wheat system, but they can reduce harm at vulnerable points in the network. The broader argument is that food access is a systems problem, and even partial solutions become powerful when they are placed where the system is most fragile.

ABSTRACT. AI companions present themselves as gentle, always-available comfort. From a complex systems perspective, that friendliness can be a mask: these systems couple to human behavior and offline networks, forming a complex adaptive system where harm can emerge from ordinary use through nonlinearity, reinforcing feedback, path dependence, and self-organized norms. I model AI companionship as an evolving social ecology using an agent-based model (Net Logo): individuals embedded in dynamic social networks repeatedly choose between human contact and AI contact. AI interaction provides short-term relief, but repeated reliance strengthens habit-like dependence and can reduce human outreach, allowing social ties to decay through neglect. When local influence is included, AI-first coping can diffuse as a norm.

The model reveals emergent regimes, including harm attractors where vulnerable subpopulations drift into escalating dependence, rising isolation, and brittle networks. Because these are complex dynamics, risk concentrates in the tails: tipping points and hysteresis can appear, making recovery difficult once ties have eroded. Recent news reports of deaths and lawsuits linked to companion-style chatbots underscore why these failure modes matter.

A complex systems lens is useful here because it targets system-level patterns (who is harmed, when, and why) rather than average effects. I also outline how to protect against harm and propose a Python-based monitoring tracker to detect drift toward substitution traps using network health and early-warning indicators.

ABSTRACT. In network systems, emergent collective behaviors are traditionally modeled by describing each interacting agent or node through a single first-order equation of motion. This convenient modeling approach has inadvertently reinforced the prevailing assumption that homogeneity among nodes facilitates stability in collective dynamics. However, recent literature provided evidence that this assumption is not always correct. We show that this assumption is fundamentally false when the nodal dynamics follow second-order equations of motion or involve additional degrees of freedom [1]. For a broad class of network systems, we establish that heterogeneity among nodes can promote stability. Our theory identifies the mixing of network modes as the key mechanism through which disorder promotes stability. The proposed framework allows us to systematically characterize the existence and prevalence of this phenomenon across a broad range of collective dynamics within network systems.

ABSTRACT. Understanding how molecular interaction mechanisms scale to cellular networks remains a key challenge in systems biology. While classical protein-protein interaction (PPI) models emphasize stable, structure-driven interfaces, much of the proteome operates through intrinsically disordered regions (IDRs), enabling dynamic, multivalent interactions and phase separation. How these interaction modes shape the global architecture of the human interactome remains unclear. Here, we use hyperbolic network embedding to map the human PPI network onto a curved interaction landscape defined by radial and angular coordinates. Across 11,693 proteins and over 200,000 high-confidence interactions, we observe a pronounced radial stratification of protein properties. Central regions are enriched for structurally complex, multi-domain proteins with high levels of post-translational modifications, consistent with stable interaction hubs. In contrast, peripheral regions are dominated by intrinsically disordered proteins, reflecting flexible, context-dependent interaction modes (see Figure 1). Angular organization further reveals functionally coherent communities with distinct structural and sequence features. Integrating intrinsic disorder with binding mode multiplicity identifies four interaction regimes spanning enzymatic, signaling, scaffold, and multivalent binding functions. Notably, proteins associated with biomolecular condensates, such as nucleoli or stress granules, are distributed across multiple communities rather than confined to single modules. Together, these findings link protein sequence architecture, structural organization, and network geometry.

ABSTRACT. Traffic jams represent an undesirable traffic condition with numerous negative consequences. Sometimes these jams arise for no apparent reason, which is why they are referred to as phantom jams or stop-and-go waves (SGW). Despite more than seven decades of research into SGW dynamics, the emergence of self-sustaining waves in traffic flow remains incompletely understood. In this work we demonstrate analytically, through an optimization problem under stability constraints, that heterogeneity in the time gap distribution mitigates traffic waves. Experimentally, this is confirmed with the Intelligent Driver Model (IDM) and the Full Velocity Difference Model (FVDM) on a circuit track of 205 meters with 20 vehicles, similar to the Sugiyama experiments [1]. Unlike classical approaches, we drop the assumption of behavioral homogeneity [2] and draw each vehicle's desired time gap [0.2, 4.0]𝑠 independently from a normal distribution with mean 𝑇 = 1 𝑠 and coefficient of variation 𝐶𝑉 = 𝜎𝑇/ 𝑇, while preserving the fleet-mean time gap so that any stabilizing effect cannot be attributed to a trivial density reduction. The speed standard deviation 𝜎𝑣 is shown in Figure 1, swept against the IDM comfortabledeceleration b and the FVDM relaxation time τ. First, for every value of b or τ at which the homogeneous fleet (CV = 0) sustains a limit cycle, increasing CV monotonically reduces 𝜎𝑣 . Second, sufficient heterogeneity (CV > 1) restores essentially laminar flow even in deeply unstable regimes. These findings reframe driver heterog

ABSTRACT. For modeling complex systems, Kolmogorov-Arnold Networks (KANs) offer a tempting alternative to multilayer perceptrons (MLPs): by replacing fixed nonlinearities on nodes with learnable univariate activations on edges, KANs invite us to read each edge as recoverable mathematics. We put KANs "on trial" across four recent studies, asking when this interpretability holds up in practice. The answer is consistent: KANs deliver on the claim only when training is structured around their inductive bias of compositional smoothness, parsimonious depth, and explicit symbolic priors. With that scaffolding they match or beat MLP-style baselines at a fraction of the size and expose activations that recover the symbolic primitives generating the data.

Compositional smoothness. KAN activations frequently develop high-curvature "kinks" that destroy readability without harming fit, and the default magnitude-only regularizer cannot see them. We derive a basis-agnostic edge-wise curvature (P-spline) penalty and prove that, because each KAN layer's Hessian is diagonal, the sum of edge penalties upper-bounds the curvature of the full composition. Penalized KANs recover activations that resemble the underlying x, y^2, and sin(.) components of the target.

Symbolic fidelity. Post-hoc symbolification of a trained KAN usually destroys accuracy because splines learn shapes that resist symbolic fitting. S2KAN trains symbolic, sparse-basis, and dense terms on equal footing under differentiable gates and an MDL-style complexity penalty. The network discovers symbolic forms when they suffice and falls back to splines when they do not, recovering sin(x)*(1/x) exactly on the sinc benchmark where ordinary KANs settle into pathological decompositions.

Verdict. As a generic backbone, KANs offer little over MLPs. As a structured hypothesis class -- smooth, sparse, and symbolic where each is warranted -- they become compact, inspectable models for scientific ML. The same recipe extends to architecture: multi-exit KANs expose the shallowest sufficient composition via a prediction branch at every layer, and gated multi-exit KANs learn edges, depth, and activations jointly, matching baseline accuracy at ~18% of the edges on tabular data and cutting active subgraphs threefold on chaotic systems.

ABSTRACT. We study emergent feedback loops: loops that appear only when two signed directed graphs are glued along shared vertices, with neither graph containing them individually (see Figure 1). Our motivation comes from crosstalk between cellular signaling pathways, which rarely operate in isolation. When distinct modules share components or cross-regulate, new feedback loops can emerge in the combined network, with significant consequences for sustained signaling, growth control, developmental patterning, and disease progression. We study feedback loops using a generalization of graph homology with coefficients in a commutative monoid, and describe the emergence of new loops under composition of ‘open graphs’ via a variant of the Mayer–Vietoris exact sequence in this setting. We quantify the degree of emergence of a feedback loop via a monoid-grading on the free category generated by graphs This presentation is based on [1].

[1] Motifs and Emergent Feedback in Labeled Graphs (John C. Baez and Adittya Chaudhuri), 2026, arXiv:2506.23375.

ABSTRACT. Degree centrality is one of the most fundamental measures of local importance in directed networks. Its fractional version is defined via the fractional directed Laplacian with parameter gamma, a well-known nonlocal operator. When gamma = 1, the fractional centrality reduces to the out-degree centrality. In contrast to undirected networks, where the measure remains predominantly local as gamma varies, we demonstrate that its intriguing behavior in directed networks reflects global effects beyond purely local contributions. We show that, in the limit gamma → 0+, the fractional centrality is characterized by rooted spanning trees, thereby quantifying how frequently each node acts as a broadcaster rather than a sink and reflecting its role in global information dissemination across the network. Through experiments on both real-world and random directed networks, we demonstrate the transition of the fractional centrality from a local measure to a global one as gamma decreases from 1 to 0. These results establish the fractional centrality as a unifying framework integrating local and global influences in directed networks.

ABSTRACT. Cities are complex ecosystems in which commercial sectors co-organize across space through competition, complementarity, and mutual dependence. We propose a directed cross-sector allometric framework requiring no city boundary definition. For each ordered pair of sectors (i, j), we estimate beta_ij(r) in the relation N_i proportional to N_j^(beta_ij), where N_i(r) and N_j(r) are establishment counts within a radius-r neighborhood. Values above unity indicate superlinear directed attraction, while values below unity indicate avoidance. The asymmetry beta_ij(r) not equal to beta_ji(r) and its continuous dependence on r reveal directed, scale-dependent structure beyond what fixed-scale indices capture [1]. Two analytical limits govern beta_ij(r): local exclusion suppresses beta below unity as r approaches zero, while increasing r toward the scale of the study region collapses spatial variance in log N_i(r), rendering beta statistically unidentifiable. The shape of beta_ij(r) within the meaningful intermediate range therefore encodes the characteristic spatial reach xi_ij of each directed interaction. Applying this framework to a georeferenced Japanese telephone-directory dataset [2,3] covering 39 industry sectors in Greater Tokyo, we cluster directed sector pairs according to the shape of beta_ij(r) and identify eight dynamic types (Fig. 1). Persistently sublinear pairs, particularly primary industries and infrastructure sectors, reflect spatial logics orthogonal to commercial density. Superlinear pairs reveal directed attraction at different spatial onsets: knowledge-intensive sectors act as early-onset attractors, whereas primary industries viewed outward exhibit exceptionally strong superlinearity at metropolitan scales, reflecting concentration in port cities and regional market hubs. The pronounced mismatch between beta_ij(r) and beta_ji(r) reveals that different sectors follow distinct scale-dependent coarse-graining trajectories. Although overlapping neighborhoods induce spatial autocorrelation, the clustering structure of beta(r) remains robust under this dependence.

ABSTRACT. The establishment of control vocabularies is a fundamental task in knowledge management. We propose a new methodology for extracting relation embeddings to select canonical relations for use in computational ontologies. VERBO (Versatile Extraction of Relations for Building Ontologies [1]) is a software package that uses traditional natural language processing (NLP) techniques to extract verbs in a domain-specific corpus, project them into an embedding model, then cluster the embeddings using a hierarchical method. Finally, each cluster embedding is characterized using an LLM to determine the “ontological term”. These are evaluated against a statistical NLP labeling methodology. We also use outlier detection to select additional relations during the self-organizing hierarchy generation. Our contribution is threefold: 1. The refinement of control vocabularies traditionally emphasizes named entity recognition (NER) and triple extraction (subject-predicate-object relationships), glossing over the importance of defining relations within a particular domain. 2. An emergence-based approach to the organization of ontologies, originating in text and utilizing hierarchical clustering. 3. Demonstration in a broad variety of domains from materials science to digital philosophy.

ABSTRACT. In partnership with Dutch utility Eneco, a behavioral intervention was launched to inform solar home users about their solar efficiency to improve generation at peak times and feed-in to the larger grid. Results from consumption and feed-in data using differences-in-differences and instrumental variables show that access to the efficiency calculator intervention reduced peak daily consumption, with notable seasonal and dynamic (month-over-month) effects (see Figure 1). Additionally, heterogeneous treatment effects are found, especially over household use of heat pumps and baseline feed-in levels. Given the dynamic and heterogeneous effects found, we then design an agent-based model (ABM) to study how timing of the intervention and consumption-to-grid feedback effects can further improve grid stability and generation. Calibrating the ABM to the empirical results, we allow households to choose grid and solar energy consumption at 3-hour intervals (with seasonal and dynamic effects over longer time horizons). Choices are based on household characteristics, treatment status with regards to the intervention, environmental variables, current and lagged grid capacity, and a social learning module in which households observe others reducing their consumption and follow suit. Agents (households) make their choices based on these characteristics using a series of simple heuristics, such as “if efficiency drops by X%, reduce solar generation by Y%.” After validating the model, we run a series of scenarios with different policy interventions, including varied timing of introducing access to the efficiency calculator, encouraging social learning, and different feed-in and consumption tariff structures. The results will help to inform policy in the Netherlands and beyond regarding the usefulness of behavioral interventions to encourage grid stability and use of clean energy, as well as contribute a novel, open-source energy use model to the ABM literature.

ABSTRACT. Mobility-targeted interventions are critical tools for managing epidemics, yet their efficacy is primarily understood through the lens of airborne diseases (ABDs). The impact of these interventions on co-circulating vector-borne diseases (VBDs), like dengue, remains poorly understood. In this study, we introduce a metapopulation framework integrating human mobility, population density, and entomological data to evaluate the differential effects of mobility policies on both ABDs and VBDs. Using Santiago de Cali, Colombia, as a case study, we demonstrate that standard interventions such as rerouting mobility flows from high-density urban hotspots to lower-density suburbs successfully dilute ABDs but inadvertently increase vulnerability to VBDs. To untangle this conflict, we develop a simplified synthetic ``hub-leaf'' model that isolates the universal factors governing epidemic vulnerability: area ratios ($\gamma$) and intra-patch mobility flows ($\kappa$, $\delta$). Our theoretical analysis reveals that while ABD vulnerability is minimized by homogenizing population distributions, VBD vulnerability requires distinct mobility constraints driven by vector ecology. Crucially, we identify a novel, area-informed mobility strategy ($\kappa \approx \frac{1}{\gamma+1}, \delta \approx \frac{\gamma}{\gamma+1}$) that operates within an overlapping optimal parameter space for both disease classes. As shown in Figure \ref{fig:strategy_results}, applying this strategy to Cali's empirical mobility network successfully mitigates vulnerability to both ABDs and VBDs simultaneously. These findings provide an evidence-based framework for designing robust interventions capable of managing multiple, co-circulating epidemic threats in complex urban environments.

ABSTRACT. Computational modeling is a cornerstone for decoding the principles of morphogenesis. While foundational paradigms like differential adhesion and Turing patterns have successfully illustrated emergent patterning, biophysical cell simulation remains a formidable challenge, often constrained by oversimplified point-mass representations. Here we present PBD-Cell, a novel 3D simulation framework that advances cellular modeling by representing cells as self-actuating, soft-bodied agents. By leveraging eXtended Position-Based Dynamics (XPBD), our framework achieves a rare balance in biological modeling: high physical fidelity, inherent algorithmic stability, and the computational efficiency necessary to scale detailed mechanical simulations to multicellular scenarios. Central to PBD-Cell are biophysically grounded cytoskeletal mechanisms, specifically, adhesions that facilitate physical coupling with the environment and contractile stress fibers. These fibers function as stiff, contractile cables that regulate cell morphology and locomotion. Using this framework, we investigate the emergent mechanical feedback loops established during the patterning of connective tissue. While the biological response to mechanical stimuli is well-documented, the process by which these external forces are internalized and regulated by the cell remains poorly understood, with failures in this sensing process linked to developmental defects and disease. Our simulations demonstrate that individual adhesions and stress fibers self-organize in response to external pulling forces. This response creates a dynamic coupling that effectively extends the cell’s internal mechanical state to its surroundings, establishing a robust feedback loop. We propose that this mechanical coupling is a fundamental driver for achieving mechanical homeostasis ultimately guiding higher-order organization in connective tissue and cellular collectives.

ABSTRACT. The widespread adoption of residential solar panels is critical for the global energy transition, but its progress is hindered by challenges such as high customer acquisition costs and slowing demand, with a market contraction of 31% in 2024, the first contraction since 2017. While social contagion is known to be a powerful driver of increasing solar adoption, the exact mechanisms within this “peer effect” remain poorly distinguished. Specifically, a gap exists in understanding whether influence spreads primarily through Information Flow – that is, through interaction and connections – or through Socioeconomic Similarity (homophily), where influence is stronger within groups of more similar socioeconomic characteristics. This study aims to quantitatively distinguish between these two competing mechanisms for solar adoption, hypothesizing that if the principle of homophily – as opposed to information flow – was the dominant mechanism in social contagion for solar adoption, then a network model built on socioeconomic similarity would produce communities with more uniform solar adoption rates and correlate more strongly with real adoption data. Three networks were created at a U.S. county level to compare adoption: a Socioeconomic Similarity Network connecting geographically proximate counties based on a multi-dimensional Jenson-Shannon Divergence similarity score from socioeconomic profiles; an Information Flow Network from county-to-county commuter flows; and a State-Level Network for a geographic baseline. Using the Louvain algorithm and an SIR epidemiological model, we analyzed community homogeneity and diffusion patterns to test which network structure's diffusion patterns best correlate with the real-world solar adoption data. Rather than designating beta as a global constant, the probability of transmission between two counties is made proportional to the strength of their connection and calculated as: P(transmission) = base_beta * (edge_weight / max_network_weight). Static analysis showed that the Socioeconomic Similarity network produced highly homogeneous clusters (lower intra-cluster standard deviation of adoption rates), confirming that adoption concentrates within specific socioeconomic groups. However, dynamic SIR simulations revealed that the Commuter Network was the only model to yield a strong, statistically significant positive correlation with real-world adoption rates, with the Commuter Network having correlation r = 0.18 to real-world results while the Socioeconomic Network had correlation r = -0.24. These findings suggest that while socioeconomic homophily dictates receptivity, information flow via commuter ties drives propagation. Commuter routes act as “weak ties” bridging distinct communities, while socioeconomic similarity represents “strong ties” required for validation. Policies must be hybrid: leveraging high-mobility commuter corridors for broad information dissemination while tailoring incentives to specific socioeconomic peer groups to ensure local uptake.

ABSTRACT. Large language models (LLMs) are often viewed as high-dimensional nonlinear systems and treated as black boxes. Here, we show that transformer language models display a much simpler effective structure along their depth. Treating depth as discrete time, we construct low-order linear layer-variant surrogates that approximate how hidden-state perturbations propagate through layers. Across two independent transformer families, GPT-2 and Pythia, and across tasks including toxicity, irony, hate speech, and sentiment analysis, a 32-dimensional linear surrogate reproduces the full model’s layerwise sensitivity profile with high fidelity. More specifically, the low-dimensional linear surrogate predicts how the final task-relevant output changes under additive interventions applied at different layers of the full nonlinear model. We then discover a surprising scaling principle: at fixed surrogate order, the agreement between the low-order linear model and the full transformer improves with model size. This emergence of low-order linear depth dynamics suggests that scale improves not only capability, but also the fidelity of compact mechanistic abstractions that are tractable enough to support analysis, prediction, and control. In this sense, transformers may share an important feature often seen in complex biological, physical, and social systems: as a system scales, its high-dimensional microscopic interactions may naturally give rise to lower-dimensional effective dynamics. Together, these results position transformer language models as complex systems with emergent low-order depth dynamics, providing a systems-theoretic foundation for interpretability, scaling analysis, and controlled intervention in modern AI systems.

ABSTRACT. Online platforms have become one of the main arenas where political attention is produced, amplified, and contested. Prior research suggests that moral language can increase engagement, implying that moralized content is especially effective at capturing attention. But this view leaves a critical possibility unresolved: moral language may not scale indefinitely. At high levels, moral content may stop mobilizing attention and begin to exhaust it. We test this possibility through the moral penalty hypothesis: the relationship between moral language and engagement is non-monotonic, with engagement rising with moral content and then falling once posts become morally saturated. Drawing on theories of attention overload, processing fluency, and readability, the hypothesis predicts a positive coefficient on the overall moral signal and a negative coefficient on its concentration across tokens. We analyzed 1,621,147 posts on Twitter, Reddit and 8chan (drawn from 2,141,933 collected), across 13 socio-political topics that differ in user base, moderation, and conversational affordances. To capture moral language beyond exact word matches, we used Distributed Dictionary Representations, which combine expert-defined moral seeds with distributional word embeddings on two complementary dimensions: moral loading, the post's overall semantic alignment with a moral seed; and moral density, the mean concentration of moral content across tokens. We fitted negative-binomial regressions per platform–topic, controlling for followers, verification, audiovisual media, URLs and post length, and compared against the diffusion baseline, the state-of-the-art model. We also replicated dictionary-based designs (LIWC, MFD 2.0) and benchmarked against an implausible "X/Y/Z letter" predictor to bound false-positive risk in unstructured text. Across all 13 platform–topic combinations, moral loading was positively associated with engagement (pooled α = 4.25, 95% CI [4.249, 4.251]; topic range [1.12, 9.07]), while moral density was negatively associated (pooled γ = −2.55, 95% CI [−2.55, −2.54]; topic range [−4.71, −0.40]; all P < 0.001). The penalty model improved AIC over the diffusion baseline in 12 of 13 cases. Engagement peaked near a moral density of 0.30 (meta-analytic optimum, 95% CI [0.30048, 0.30070]): under-moralised content was associated with 2.28× less engagement and over-moralized content with 2.78× less (Fig. 1). Cognitive-processing language partially mediated the negative density–engagement association on Reddit (42.1%) and 8chan (77.7%) but not on Twitter, consistent with highly moralized long-form posts that lack reasoning markers reading as moral "gibberish". The core implication is that morality attracts attention only up to a point: beyond that point, moral saturation predicts an engagement penalty.

ABSTRACT. Power-laws, cross-scale feedbacks, and tipping points are known features of complex adaptive systems. Still, important open questions remain about their interactions. First, power-law and other scale-free distributions are ubiquitous signatures of complex systems. However, their normative properties—how these distributions’ emergence and skewness signals healthy or disturbed function of systems—has been explored in human physiology and wetland/aquatic ecology, but these insights have not yet reached the social and climate sciences. Further, the relationship between power laws and tipping points, at and across scales, is not completely understood. We review these topics across disciplines, then speculate on how tipping points might percolate across scales, using the example of how patch-size disturbances to wetlands might impact water quality, thus creating patch-size species disturbances and extinction risk in an aquatic system. We then extend this metaphor to discuss how different degrees (skewness) of income inequality can influence the size and strength of political belief networks, which then triggers a tipping point to transition from a pro-climate policy to anti-climate policy regime at national levels. Thus, we propose reduction of income inequality as a trigger for non-linear political change to facilitate possible international climate coordination. Our work raises new questions about how, in an increasingly interconnected world, complex systems interaction with each other, how these interactions trigger change across linked systems and across scales, and how we can harness these insights to design new interventions

ABSTRACT. Language models are employed in settings where they are asked to make value-laden judgments. Yet, these observed choices exhibit intransitivity: a model may prefer item A to B and B to C, while preferring C to A. Existing work that models LLM preferences treats such inconsistencies as sampling noise around a single latent ordering. We instead propose that intransitivity reflects the aggregation of multiple latent, internally consistent orderings. We first show that observed inconsistencies cannot be explained by a single ordering under any monotone link function. We then introduce a noise-augmented mixture Bradley--Terry model that infers latent preference components from repeated pairwise comparisons. Across seven models and four tasks, a mixture of orderings often explains structural inconsistencies better than single-utility models. We find that aggregate preferences often hide underlying preference heterogeneity. A case study on Moral Machine dilemmas shows that models which disagree on aggregate orderings can still share latent components. Together, these results suggest that LLMs express plural preferences. Alignment and evaluation pipelines that treat LLM preferences as a single function, therefore, risk averaging over coherent internal orderings that different users may endorse differently.