Co-Chair
Co-Chair
Photrek: Megan Hess and Igor Oliveira
Active Inference Institute: Daniel Friedman, Alexandra Mikhailova, Alexander Sabine
Karl Friston, University College of London;
Ugur Tirnakli, Izmir Economics University;
Johan Medrano, Massachusetts Institute of Technology
The 2026 International Conference on Complex Systems (CCS 2026) offers a premier forum for addressing the dynamics of interconnected systems. The Thriving the Extremes satellite session, co-hosted by Dr. Kenric Nelson and Dr. André L. M. Vilela, addresses a critical boundary condition in cognitive modeling, artificial intelligence, and statistical physics: the survival of agents in non-equilibrium environments. In non-equilibrium, the assumption of a conditionally independent boundary, called the Markov Blanket, between an agent's internal states and external environment is violated, requiring new analytical methods.
At the core of contemporary cognitive modeling is Active Inference, which posits that self-organizing systems sustain their integrity by minimizing variational free energy. A fundamental mathematical requirement for this behavior is the Markov blanket — a statistical boundary, or a set of variables, that assures conditional independence of all other variables in the system, such as an agent's internal states and the external environment. In standard thermodynamic formulations, the entropy scales linearly, and extreme outliers decay rapidly. However, real-world complex adaptive systems, such as turbulent financial markets and deeply polarized social networks, exhibit scale-free topologies and synchronized critical fluctuations driven by long-range, nonlinear correlations. When a standard Active Inference agent is deployed in such a turbulent environment, its predictive coding algorithms are vulnerable to disruptions. The strict enforcement of conditional independence forces the agent to treat highly correlated, extreme systemic shifts as unpredictable noise. Consequently, the agent's Markov blanket dissolves, leading to systemic collapse.
To overcome these limitations, this session will discuss a groundbreaking theoretical architecture: a generalized free energy that separates the modeling of fluctuations from the learning of short-range correlations. Developed from the foundations of nonlinear statistical coupling, the Coupled Free Energy (CFE) framework provides the curved information geometry required to build resilient generative models. The Coupled Free Energy utilizes the unique properties of the Coupled Entropy, which, for a positive nonlinear shape (κ), balances higher penalties for rare events with an expectation over a modified distribution with faster decaying tails. The approach guarantees the ability to model, train, and infer over the most extreme heavy-tailed distributions. In AI applications, such as the Coupled Variational Autoencoder, this allows for the creation of "risk-aware" artificial intelligence capable of mitigating catastrophic outliers and maintaining stability during black-swan events.
Applying this generalized free energy framework to cognitive models yields the second major theme of the session: the generalization of the Markov Blanket for Active Inference. We propose the conceptualization and deployment of the "Coupled Markov Blanket." Unlike its traditional counterpart, the Coupled Markov Blanket abandons the strict requirement of conditional independence. Instead, it utilizes the (κ) nonlinear shape parameter to structure a boundary in which the factorization utilizes a nonlinear function modeling the long-range dependence. This topological innovation allows an agent to anticipate and mitigate heavy-tailed extremes.
Rigorous evaluation of these advanced agents requires highly volatile experimental arenas. Thus, the third pillar of the session focuses on simulating agent-based socioeconomic systems. The session will demonstrate the performance of Coupled Variational Inference agents embedded within the majority-vote model, the chaotic deterministic model, and related systems. By introducing variables such as rewiring probabilities for small-world networks, visibility parameters to model algorithmic filter bubbles, and "strong actors" representing unyielding zealots, these models generate a deeply interconnected, heavy-tailed environment.
By moving beyond linear, independent assumptions, this session bridges nonlinear statistical physics, machine learning, cognition, and sociophysics to provide the complexity community with the theoretical tools necessary to understand how artificial agents can learn, act, and thrive in the extremes.
Session 1: Foundations and Limitations of Active Inference
“Foundations and Limitations of Active Inference,” Johan Medrano and Karl Friston
“Introduction to Active Inference”, Daniel Friedman
“Predictive Coding”, Alexander Ororbia and Faeze Habibi
Q&A and Panel
Session 2: Nonequilibrium Complex Systems
“Coupled Entropy and Extreme Value Statistics,” Amenah AL-Najafi
“Modeling Socioeconomic Systems with Majority-Vote”, Andre Vilela and Bernardo Zubillaga
“Modeling Chaotic Determinism,” Ugur Tirnakli
TBD, Alexander Sabine
Q&A and Panel
Session 3: Active Inference in Non-equilibrium
“Informational Thermodynamics for Nonequilibrium, Complex Systems,” Kenric Nelson
“Active Inference in Non-equilibrium,” Igor Oliveira and Amenah AL Najafi
Please submit an extended abstract regarding one of the listed or closely related topics. Abstract should be less than 1,500 characters and may include one figure and up to 20 references.
Submit your abstracts HERE
Early Registration: August 20, 2026
To join the satellite event, registration for the CCS2025 main conference is mandatory, either for the full conference or for satellite events only. Make sure you register at: CSS2026 Registration