Selected work

Representative Papers

Coalition Free Energy and Adaptive Precision in Multi-Agent Cooperation

May 25, 2026

This work addresses the challenge of credit assignment under uncertainty in multi-agent cooperation by proposing a variational framework grounded in the Game-Theoretic Free Energy Principle (GT-FEP). The approach models agent coalitions via Gibbs distributions and integrates Shapley values with variational inference. Its key innovation lies in uncovering a non-monotonic relationship between Shapley values and perceptual precision, leading to the design of an Adaptive Precision Control (APC) mechanism that dynamically optimizes observation precision without requiring prior hyperparameter tuning. Empirical evaluations on real-world Swiss roundabout trajectory data and multi-agent control tasks demonstrate that APC adapts online to varying noise levels, achieving performance comparable to the best fixed-precision baselines while eliminating the need for laborious hyperparameter pre-tuning.

0 citationsRead paper

A Collective Variational Principle Unifying Bayesian Inference, Game Theory, and Thermodynamics

Apr 30, 2026

This study addresses the absence of a unified theoretical framework for explaining the emergence of collective intelligence in decentralized multi-agent systems. The authors propose a game-theoretic free energy principle, establishing the first variational framework that unifies Bayesian inference, game theory, and thermodynamics. By embedding local free energy minimization within stochastic games under constraints of bounded rationality and partial observability, they demonstrate that stationary points of collective free energy correspond to approximate Nash equilibria of the induced game and provide a Gibbs-distribution-based variational representation for cooperative games. The work introduces a novel free-energy formulation of the Harsanyi dividend to capture irreducible synergistic effects and reveals a non-monotonic relationship between perceptual precision and agent influence. The theory is validated across neural, biological, and artificial multi-agent systems, thereby unifying foundational principles of inference, thermodynamics, and game-theoretic equilibrium.

0 citationsRead paper

NeuroGame Transformer: Gibbs-Inspired Attention Driven by Game Theory and Statistical Physics

Mar 19, 2026

Standard Transformers struggle to capture high-order dependencies among tokens due to their pairwise attention mechanism. This work proposes a novel attention framework that integrates cooperative game theory and statistical physics: external fields are constructed using Shapley values and Banzhaf indices, system energy is defined via an Ising Hamiltonian, and attention weights are derived from the marginal probabilities of a Gibbs distribution, efficiently approximated through mean-field theory. A learnable interpolation parameter is introduced to balance fairness and sensitivity, while a Gibbs-weighted importance sampling strategy mitigates the computational burden of the exponential coalition space. The model achieves 86.6% validation accuracy on SNLI, outperforming ALBERT-Base, matching RoBERTa-Base, and significantly surpassing various efficient Transformer baselines.

0 citationsRead paper

Recent publications

Latest Papers

Coalition Free Energy and Adaptive Precision in Multi-Agent Cooperation

May 25, 2026

This work addresses the challenge of credit assignment under uncertainty in multi-agent cooperation by proposing a variational framework grounded in the Game-Theoretic Free Energy Principle (GT-FEP). The approach models agent coalitions via Gibbs distributions and integrates Shapley values with variational inference. Its key innovation lies in uncovering a non-monotonic relationship between Shapley values and perceptual precision, leading to the design of an Adaptive Precision Control (APC) mechanism that dynamically optimizes observation precision without requiring prior hyperparameter tuning. Empirical evaluations on real-world Swiss roundabout trajectory data and multi-agent control tasks demonstrate that APC adapts online to varying noise levels, achieving performance comparable to the best fixed-precision baselines while eliminating the need for laborious hyperparameter pre-tuning.

0 citationsRead paper

A Collective Variational Principle Unifying Bayesian Inference, Game Theory, and Thermodynamics

Apr 30, 2026

This study addresses the absence of a unified theoretical framework for explaining the emergence of collective intelligence in decentralized multi-agent systems. The authors propose a game-theoretic free energy principle, establishing the first variational framework that unifies Bayesian inference, game theory, and thermodynamics. By embedding local free energy minimization within stochastic games under constraints of bounded rationality and partial observability, they demonstrate that stationary points of collective free energy correspond to approximate Nash equilibria of the induced game and provide a Gibbs-distribution-based variational representation for cooperative games. The work introduces a novel free-energy formulation of the Harsanyi dividend to capture irreducible synergistic effects and reveals a non-monotonic relationship between perceptual precision and agent influence. The theory is validated across neural, biological, and artificial multi-agent systems, thereby unifying foundational principles of inference, thermodynamics, and game-theoretic equilibrium.

0 citationsRead paper

NeuroGame Transformer: Gibbs-Inspired Attention Driven by Game Theory and Statistical Physics

Mar 19, 2026

Standard Transformers struggle to capture high-order dependencies among tokens due to their pairwise attention mechanism. This work proposes a novel attention framework that integrates cooperative game theory and statistical physics: external fields are constructed using Shapley values and Banzhaf indices, system energy is defined via an Ising Hamiltonian, and attention weights are derived from the marginal probabilities of a Gibbs distribution, efficiently approximated through mean-field theory. A learnable interpolation parameter is introduced to balance fairness and sensitivity, while a Gibbs-weighted importance sampling strategy mitigates the computational burden of the exponential coalition space. The model achieves 86.6% validation accuracy on SNLI, outperforming ALBERT-Base, matching RoBERTa-Base, and significantly surpassing various efficient Transformer baselines.

0 citationsRead paper