ComplexityPod
« ComplexityPod

#23 — AI predicts tipping points, climate signal limits, agentic LLMs

September 15, 2026

Other Developments

Sources

  1. Early warning signals of tipping points: A deep learning approach to direction and timing
    The crossing of climate tipping points poses serious risk to life on Earth due to the impacts they cause, and their irreversibility. These systems generally exhibit critical slowing down on their approach to tipping point, or bifurcation, which can be detected by looking for changes in statistical properties of a time series from the system, such as an increasing lag-1 autocorrelation or variance. Recently, there has been interest in the use of deep learning in tipping point detection. However, there has been a strong focus on predicting the movement towards tipping only. Here, we present…
  2. Predicting tipping points: The many shades of non-equilibrium and catch-22s of early-warning
    The potential of crossing climate tipping points (TP) has reached the attention of many researchers and the general public. On the one hand, the basis for this concern is strengthening, with simulations showing that abrupt transitions might occur even for moderate emission scenarios. On the other hand, our understanding of what constitutes such transitions mathematically is becoming more nuanced. This leads to challenges for the fidelity of early-warning signals (EWS), which accompany bifurcations in systems that closely track a slowly changing steady state. Different kinds of…
  3. Beyond static responses: multi-agent LLM systems as a new paradigm for social science research
    Abstract As large language models (LLMs) transition from static tools to fully agentic systems, their potential for transforming social science research is well recognized. This paper introduces a structured framework for understanding the diverse applications of agentic LLM systems, ranging from simple data processors to complex, multi-agent systems capable of simulating emergent social dynamics. By mapping this developmental continuum across six levels, the paper clarifies the technical and methodological boundaries between different agentic architectures, surveying current capabilities and…
  4. The SNA-Evaluation Framework: Advancing the Use of Social Network Analysis in Program Evaluation
    As evaluation increasingly addresses the complexities inherent in dynamic program systems, social network analysis (SNA) offers significant opportunities alongside notable challenges. While SNA can reveal patterns of collaboration, communication, and influence, its application is often superficial and insufficiently integrated with evaluation theory. The Social Network Analysis for Evaluation (SNA-E) framework is introduced to address this disconnect by embedding network thinking within established evaluation traditions. This framework enables evaluators to systematically examine stakeholder…
  5. Systemic Thought in Psychology and Family Therapy: Interdisciplinary Origins, Flexibility, And African Relational Perspectives
    Objective : To reconstruct selected interdisciplinary origins of systemic thought in psychology and family therapy, clarify the movement from first- to second-order systems thinking, and examine how African relational scholarship and critiques of power revise contemporary systemic practice. Method: A critical narrative review was conducted using purposive searches of PsycINFO, Scopus, Google Scholar, Crossref, PubMed where clinically relevant, and publisher databases for foundational and interpretive literature on general system theory, cybernetics, family therapy, feminist systemic critique,…
  6. The pathology of trivialization: bureaucracy as a “first-order machine” against “non-trivial” citizens
    Purpose This theoretical review examines a fundamental epistemological error in public administration: the tendency of the state to act as a “first-order external observer”. From this perspective, the state apparatus designs policies by viewing society as a predictable “Trivial Machine”, expecting perfectly linear outcomes. Observations, however, suggest that citizens operate as highly complex systems equipped with memory and the capacity for strategic adaptation. Design/methodology/approach The present analysis identifies one overarching theoretical gap in the literature, which unfolds…
  7. Adaptive Cycles in Open Innovation: A Longitudinal Study of Circular Economy Collaborations in a Fashion Incumbent
    ABSTRACT This single case study examines how the H&M Group, a large fashion multinational, uses open innovation (OI) to support the transition towards a circular business model. Although collaborations are recognized as key to advancing circularity, limited empirical evidence exists on how co‐creative OI processes drive circular practices across the value chain. Using adaptive cycle theory (the panarchy model) and institutional theory, the study analyses H&M Group's involvement in circular and collaborative innovation from 2010 to 2025 through a mixed‐methods approach, including case study…
  8. Early warning signals of sudden changes in daily step count: an intensive longitudinal study of over 20,000 adults
    Abstract Health-protecting behaviors such as physical activity are crucial for health, yet positive behavior changes are often accompanied by setbacks and relapses. Previous research suggests that critical fluctuations in behavioral time series might foreshadow relapses as early warning signals. This study examines the association between critical fluctuations, quantified as dynamic complexity (DC), and impending sudden changes in physical activity, here walking behavior. We analyzed step count data from 21,425 participants during enrollment in a digital physical activity intervention,…
  9. Early warning signals for synchronization transitions from partial observations
    Anticipating the onset of collective synchronization is important in many networked systems, yet observing every oscillator is often impractical. We investigate whether synchronization transitions can be detected from a small set of monitored, or sentinel, nodes. Using a stochastic Kuramoto model on networks, we numerically compare three early warning signals: the local order parameter, its temporal variance, and the variance of individual oscillator phases after removing their mean rotational trends. We also compare sentinel-selection strategies based on node dynamics, degree, and random…
  10. A Temporal Multiplex Graph Neural Network for Systemic Risk Transmission in Global Banking
    This paper develops a unified framework for assessing systemic risk and identifying contagion channels in the global banking system using a Temporal Heterogeneous Multiplex Graph Neural Network. We construct a harmonised quarterly panel combining bank fundamentals, CDS spreads, and macroeconomic indicators, and represent these data as dynamic multiplex networks linking banks through financial similarity and liquidity co-movement, augmented with country-level macroeconomic relationships. The model integrates graph convolutional layers with recurrent GRU dynamics and incorporates a learnable…
  11. Increasing school attendance and retention in refugee settings: Participatory development of the SchoolLinks intervention
    This study aimed to develop an intervention to improve consistent school attendance in participation with caregivers, teachers, and other stakeholders, in refugee settings in Uganda. In Phase 1, we organized four group model building workshops with caregivers and teachers in four refugee settlements to identify factors influencing attendance and caregiver engagement in education, and generate proposed actions to address them. We synthesized the outputs into a causal map and set of intervention strategies. These were reviewed through four respondent validation workshops with caregivers,…

Also this week

Full transcript
Neural networks can now spot the structural warning signs of collapse before a complex system crosses the threshold. That capacity to forecast critical transitions is where we begin on ComplexityPod, examining systems across research domains. Here are the latest papers. Detecting critical transitions in dynamical systems usually relies on one core assumption: critical slowing down. We watch for rising variance or lag-one autocorrelation. Which works when a system is drifting toward instability, but those indicators do not tell you whether you are moving toward a threshold or backing away from it. A study from August 24 addresses that limitation directly. The researchers paired convolutional neural networks with gated recurrent units to classify trajectories relative to critical boundaries. And they trained it on three bifurcation normal forms with linear forcing. That setup let them test both directions—approaching a bifurcation and retreating from one. Right, so instead of just flagging instability, the architecture determines trajectory direction and estimates time to the threshold. It hit seventy-eight percent accuracy on extended series. They also ran it against climate simulations—specifically Atlantic Meridional Overturning Circulation collapse in general circulation models. It separated the forced runs from unforced baselines and mapped the changing distance to the collapse point. Though on the climate side, another paper from September 9 raises a structural hurdle to that entire category of detection. The issue of non-equilibrium states. Exactly. Critical slowing down assumes adiabatic conditions—a system tracking slowly shifting steady states. But rapid forcing, multistability, and chaos pull high-dimensional systems out of equilibrium entirely. Which means local stability indicators lose their footing. If a transition happens without staying near a local equilibrium, local metrics cannot map the boundary. The authors argue the math has to shift toward non-equilibrium statistical mechanics and global dynamical systems. Early warning signals also surfaced in human behavioral data on August 25. An analysis of over twenty-one thousand adults across five million person-days examined step counts in a digital health program. Using dynamic complexity to capture fluctuations before sudden behavioral shifts. And the signal was directional. Higher dynamic complexity preceded sudden drops in physical activity, with odds ratios between 1.32 and 1.93. Yet it showed weak and inconsistent correlations with sudden gains. The instability metric basically served as an early marker of activity loss, not sudden increases. Monitoring instability in networks faces a separate challenge: you cannot always measure every node. An August 28 paper looked at stochastic Kuramoto oscillator networks undergoing synchronization transitions. Testing sentinel node selection instead of whole-network monitoring. They checked nodes chosen by dynamical behavior, network degree, and random sampling. Dynamic selection performed on par with monitoring the entire network, and the required sentinel sample size scaled only logarithmically with total network size. What stood out was the choice of metric. The local order parameter and the variance of detrended phases tracked the transition far better than local order parameter variance. Tracking state shifts across networks also appeared in finance on August 27. A study introduced a temporal heterogeneous multiplex graph neural network to measure systemic risk in global banking. Multiplex networks link institutions across multiple layers at once—here, financial similarity and liquidity co-movement, tied to macroeconomic factors. The architecture combined graph convolutional layers, recurrent gated units, and a learnable fusion gate to observe contagion channels as they shifted. In stress tests and credit default swap spread forecasting, it beat standard econometric baselines. Moving from financial networks to institutional networks, an August 22 paper introduced SNA-E, a framework connecting social network analysis to program evaluation. It organizes network methods into four domains: network constructs, evaluation types, evaluation purposes, and network metrics. The intent is to give evaluators a systematic process for tracking stakeholder coordination and resource flows in multi-agency projects. Supply chains saw a similar structural lens on August 31. A longitudinal study of the H and M Group between 2010 and 2025 applied panarchy adaptive cycle theory and institutional theory to circular economy shifts. Tracing how open innovation efforts moved through all four adaptive cycle phases across seven collaboration categories, shaped by rules, financial constraints, and market pressures. Administrative systems received a cybernetic critique in a September 5 paper on public governance failures. The authors reviewed 128 papers and identified what they call the pathology of trivialization. Where an agency operates as a first-order observer, treating citizens as trivial machines with direct linear inputs and outputs. While citizens actually operate as non-trivial systems with memory and adaptive responses. That mismatch triggers a four-stage loop: Bureaucratic Trivialization, Non-Trivial Adaptation, Administrative Debt Accumulation, and finally Reflexive Governance. Their proposed fix requires second-order cybernetic feedback—institutions altering their internal rules rather than offloading rule-bending onto frontline worker discretion. Modeling complex social feedback showed up in social science methodology on September 2, with a six-level taxonomy for agentic large language models. It scales from simple text annotation up to autonomous multi-agent systems simulating group norms and cultural coordination. But the paper also maps the hard boundaries: validation, reproducibility, and emergent biases in multi-agent environments. Cybernetics also crossed into clinical practice on August 26. A narrative review developed a three-level rigidity-flexibility framework for systemic psychology. Tracing the field from first-order models to second-order reflexivity, while incorporating Ubuntu relational philosophy and power analyses. It questions complete therapeutic neutrality, arguing therapists must intervene directly against coercive control while supporting agency in family structures. Community-level systems work appeared in Uganda on August 31. Group model building workshops in four refugee settlements mapped retention barriers with teachers, caregivers, and administrators. They mapped interactions among caregiver encouragement, language barriers, teacher communication, and economic pressures, producing SchoolLinks—a set of six collaborative strategies for student retention. Several other papers completed this cycle's radar. Digital engineering transformations in complex engineered systems sustainment demonstrated measurable schedule reductions. A narrative review examined intensive care unit operations through complex adaptive systems theory. A synthesis of 197 studies evaluated social-ecological resilience across Karst World Natural Heritage buffer zones, and a study in the Yiluo River Basin paired synergetics with random forest models for socio-hydrological adaptability. Botanists documented a time-dependent bistable switch regulating floral transitions in Arabidopsis. A network analysis evaluated systemic risk transmission between Morocco's banking sector and its national economy. And a bibliometric study of bank interconnectedness from 2013 to 2025 tracked a clear structural transition in the literature toward network topology and adaptive supervision. We will return next week with another collection of research across complex systems. From ComplexityPod, thanks for listening.

From our community

At the Intersections — latest issue

Welcome to this issue of At the Intersections. This edition brings together observations across environmental policy, neurological function, and human social experience. The selections connect discussions surrounding life in the soil with c

Read the latest issue →