A Collective Variational Principle Unifying Bayesian Inference, Game Theory, and Thermodynamics
arXiv:2604.27942v1 [cs.AI] 30 Apr 2026
Djamel Bouchaffra1,* , Faycal Ykhlef2 , Mustapha Lebbah1 , Hanane Azzag3 1 DAVID Lab, University of Paris-Saclay, UVSQ, 78035 Versailles, France. 2 TSAM, ASM, Centre for Development of Advanced Technologies, Algiers, Algeria. 3 LIPN, UMR CNRS 7030, Sorbonne Paris Nord University, Villetaneuse, France. *Correspondence: [email protected] Abstract Collective intelligence emerges across biological, physical, and artificial systems without central coordination, yet a unifying principle governing such behaviour remains elusive. The Free Energy Principle explains how individual agents adapt through variational inference, while game theory formalises strategic interactions. Here we introduce the Game-Theoretic Free Energy Principle, a unified framework showing that multi-agent systems performing local free-energy minimisation implicitly implement a stochastic game. We prove that, under bounded rationality and local information constraints, stationary points of collective free energy correspond to approximate Nash equilibria of an induced game. Conversely, a broad class of cooperative games admits a variational representation in which equilibria arise as Gibbs distributions over coalitions, establishing a bridge between Bayesian inference and strategic interaction. To characterise higher-order effects, we introduce a free-energy formulation of the Harsanyi dividend, isolating irreducible multi-agent synergy. This yields a predictive theory of cooperation, including a falsifiable non-monotonic relationship between sensory precision and agent influence. We validate this prediction across neural, biological, and artificial multi-agent systems. These results identify a common variational principle underlying inference, thermodynamics, and game-theoretic equilibrium.
Introduction How do large populations of interacting units—from physical particles and molecules [1], and cells, to neurons, animals, and engineered multi-agent systems-achieve coherent collective behaviour without centralised control? This question arises across neuroscience, biology, physics, economics, and artificial intelligence, where systems composed of bounded, locally informed agents nevertheless exhibit global coordination and adaptive structure. Recent perspectives highlight that collective intelligence spans natural and artificial systems, with common functional challenges across animals and robots [2]. A central challenge is that existing theoretical frameworks address complementary but incomplete aspects of this phenomenon. The Free Energy Principle provides a general account of how individual systems maintain adaptive organisation by minimising variational free energy through Bayesian inference [3, 4, 5, 6]. However, this formulation is inherently single-agent and does not explain how multiple agents coordinate, compete, or form synergistic coalitions. Recent efforts to extend the Free Energy Principle to multi-agent systems have explored how groups of agents can share beliefs and world models to achieve a common understanding [?], how they can form a larger, group-level agent through a collective Markov blanket [?], and how robust decision-making in multi-agent settings might be achieved [7]. However, these approaches focus on emergent communication and group-level inference, rather than providing a formal game-theoretic analysis of strategic interactions, a computable measure of irreducible synergy, or falsifiable predictions about individual influence within a coalition. In parallel, game theory offers a descriptive and normative theory of strategic 1
interaction, with equilibrium concepts such as Nash equilibria [8, 9] and tools such as coalitional decompositions [10]. Recent work has begun to bridge game theory with statistical physics in deep neural network architectures [11, 12]. Coalitional game theory has also been used to interpret deep neural networks [13], underscoring the relevance of such tools in understanding complex systems. Yet classical game theory lacks a mechanistic grounding in inference or physical principles, and does not naturally explain how equilibrium behaviour emerges from local probabilistic computation in distributed systems. Here we propose a unifying perspective in which multi-agent systems are described as performing distributed variational inference over joint configurations induced by their shared environment. In this framework, each agent minimises its own variational free energy, and interactions among agents induce an implicit stochastic game over coalition structures. We show that, under bounded rationality, stochastic policy selection, and local information constraints, stationary points of the resulting variational dynamics correspond to ϵ-Nash equilibria of the induced game. Conversely, a broad class of cooperative games admits a variational representation in which equilibrium strategies arise as Gibbs distributions over coalitions. To move beyond equilibrium characterisation and capture the internal structure of cooperation, we introduce a variational formulation of the Harsanyi decomposition [14, 15, 16]. This allows the energy of a coalition to be expressed in terms of irreducible higher-order contributions, providing a direct measure of synergy and conflict in terms of free-energy reduction. Recent work on higher-order topological dynamics further supports the necessity of going beyond pairwise interactions in complex systems [17]. This perspective yields a quantitative and falsifiable prediction: an agent’s influence within a collective depends non-monotonically on its sensory precision. The emergence of cooperation in multi-agent systems has been identified as a fundamental statistical physics problem [18]. Our framework addresses this problem by linking free energy minimisation to game-theoretic equilibrium, leading to the following prediction: moderate increases in precision enhance coordination and global influence, whereas excessive precision leads to over-specialisation and reduced systemic impact due to amplification of local noise. We test this prediction across neural, biological, and artificial multi-agent systems. Finally, we show that classical models from statistical physics and machine learning, including Ising models, Boltzmann machines, and attention-based architectures [19], arise as special cases of the proposed variational framework under appropriate restrictions on interaction structure and mean-field approximations.
Results The Game-Theoretic Free Energy Principle (Theory) Axiom 1 (Multi-agent system as a variational game). A system consists of N interacting agents N = {1, . . . , N }. Each agent i maintains a generative model of its local observations õi and latent states s̃i , and updates its beliefs by minimising an individual variational free energy. Let qi (s̃i ) denote the agent’s variational approximation to the true posterior p(s̃i | õi ). Then the free energy is defined as Fi = Eqi (s̃i ) ln qi (s̃i ) − ln p(õi , s̃i ) . Agents interact through a shared environment that couples their generative models. The negative free energy defines their effective utility, inducing a stochastic game in which each agent acts under bounded rationality and local information constraints. Any subset C ⊆ N defines a coalition corresponding to coordinated inference and action. Axiom 2 (Energy of coalitions via Harsanyi decomposition of synergy). The interaction structure of a coalition C is encoded by an energy functional E(C) decomposed into
2
contributions of increasing order: E(C) =
X
ϕi +
i∈C
X
ψij +
i,j∈C
X
χijk + · · ·
i,j,k∈C
where higher-order terms arise from the Harsanyi decomposition of coalition synergy, isolating irreducible interactions that cannot be expressed as combinations of lower-order effects. For each coalition B ⊆ N , define its Harsanyi dividend ∆(B) as the unique quantity such that for every coalition C, X E(C) = ∆(B). B⊆C
This decomposition uniquely partitions the energy into additive and synergistic components. The Harsanyi dividends are computed via Möbius inversion (see Supplementary Information). Axiom 3 (Stochastic inference and bounded rationality). Agent decisions are stochastic due to noise, uncertainty, and computational constraints. This is modelled by an inverse temperature β > 0 controlling rationality. The system induces a probability distribution P (C) over coalitions, representing uncertainty over latent interaction structures. Definition (Collective variational free energy). tion distributions is defined as: F(P ) = EP [E(C)] −
The system-level free energy over coali1 H(P ), β
where H(P ) is the Shannon entropy over coalition configurations. This defines a variational inference problem over latent interaction structures. Theorem 1 (Gibbs equilibrium as variational posterior). imised by the unique Gibbs distribution P ∗ (C) =
1 exp(−βE(C)), Z
Z=
X
The functional F(P ) is min-
exp(−βE(C)).
C
This distribution defines a variational Bayesian posterior over coalition structures. (Proof in Supplementary Information.) Theorem 2 (Variational Nash–FEP correspondence). Consider a multi-agent system in which each agent minimises its individual free energy Fi under Axiom 1, and coalition interactions are encoded by E(C) under Axiom 2. Then: 1. The Gibbs equilibrium P ∗ (C) induces an ϵ-Nash equilibrium of the associated stochastic game, in which each agent cannot unilaterally reduce its expected free energy given the strategies of others. 2. Conversely, any cooperative game with characteristic function v admits a variational representation in terms of an energy function E(C) constructed from Harsanyi dividends, such that its Nash equilibrium coincides with the Gibbs posterior P ∗ . (Proof in Supplementary Information.)
3
Corollary (Synergistic free-energy structure). The Harsanyi dividend ∆(C) quantifies the irreducible contribution of a coalition to collective free energy: • ∆(C) > 0: synergistic reduction of free energy (emergent cooperation), • ∆(C) < 0: antagonistic interaction increasing collective free energy. These quantities define a principled decomposition of interaction structure that is computable from generative models without behavioural observation. Interpretation. This framework establishes a variational equivalence between Bayesian inference, stochastic game theory, and statistical physics. Agents act as local variational optimisers, while global behaviour emerges as a Gibbs posterior over coalition structures. The Harsanyi decomposition provides the interaction geometry underlying this posterior, and classical models such as Ising systems, Boltzmann machines, and attention-based architectures arise as limiting cases of the energy functional under restricted interaction structure or mean-field approximations.
Coalitional Synergy and the Harsanyi Dividend of Free Energy P From Axiom 2, the energy of a coalition C decomposes as E(C) = B⊆C ∆(B), where ∆(B) are the Harsanyi dividends. The coefficients ϕi , ψij , χijk , . . . are precisely the Harsanyi dividends of the corresponding subsets: ϕi = ∆({i}),
ψij = ∆({i, j}),
χijk = ∆({i, j, k}),
etc.
The Harsanyi dividend ∆(B) for a coalition B is computed from the coalition value function v(C) (or equivalently from the energy E(C) = −v(C)) using the Möbius inversion formula over the subset lattice. Given a value function v : 2N → R with v(∅) = 0, the Harsanyi dividend of B is: X ∆(B) = (−1)|B|−|A| v(A), A⊆B
where A is a dummy summation variable that runs over all subsets of B (including B itself and the empty set). Using the energy E(C), one can write: X ∆(B) = − (−1)|B|−|A| E(A). A⊆B
For each coalition A, E(A) is the variational free energy obtained when the agents in A jointly optimise their beliefs. It can be computed directly from the agents’ generative models and variational distributions, without any knowledge of the dividends ∆. Moreover, the sum runs over all subsets A of B (including B itself and the empty set), and for each suchP A, E(A) is already known. Once all ∆(B) are obtained in this way, the decomposition E(C) = B⊆C ∆(B) holds as a consequence of Möbius inversion, not as a definition (see Supplementary Information for a detailed derivation). From Harsanyi dividends to causal influence. The Shapley value ηi of an agent i quantifies its average marginal contribution to all coalitions [20, 21]. In terms of the Harsanyi dividends ∆(B), the Shapley value can be expressed as a weighted sum over all coalitions that contain i: ηi =
X 1 ∆(B). |B|
B⊆N i∈B
4
This formula shows that an agent’s causal influence is a linear combination of the irreducible synergies (positive ∆) and conflicts (negative ∆) of all coalitions in which it participates. Consequently, the non-monotonic behaviour of ϕi as a function of sensory precision β directly reflects the change in sign and magnitude of the Harsanyi dividends with β (see the Empirical Validation section below). In the fish schooling model [22], the influence is computed directly as the Shapley value derived from the coalition free energies, and therefore also inherits this decomposition. Similarly, the Shapley values in multi-agent reinforcement learning are computed from the same principle.
Falsifiable Prediction: Non-monotonic Influence with Sensory Precision We predict that an agent’s causal influence within a collective (measured by its Shapley value or marginal coalition probability) follows an inverted-U shape as a function of its sensory precision. At low precision, inference is poor and influence low; at intermediate precision, coordination peaks; at high precision, overfitting to local noise reduces influence. This non-monotonic law is a direct consequence of the bias-variance trade-off in variational inference under bounded rationality [23, 24, 25, 26, 27].
Empirical Validation Across Three Domains We tested the predicted non-monotonic relationship between sensory precision and individual influence in three distinct multi-agent systems spanning neuroscience, biology, and artificial intelligence. In each domain, we varied the precision (inverse variance) of each agent’s observations, measured the agent’s causal influence on collective behaviour (quantified via Shapley values or marginal coalition probabilities derived from the Harsanyi dividend), and observed the predicted inverted-U shape. Neural ensembles. Persistent neural activity is a hallmark of working memory [28]. We simulated a population of N = 50 agents (neurons) using a Gaussian generative model where each agent has a hidden state si with prior N (0, 1) and observation noise precision β (see Supplementary Information). Each agent minimises its variational free energy Fi (Axiom 1). For each precision value β in 35 steps from 0.25 to 2.0, we computed coalition free energies E(C) analytically and derived the Shapley value ηi for each agent using the exact formula for symmetric agents. The mean influence (Shapley value averaged over all agents and 50 independent runs) is reported. The results (Figure 1) show a clear non-monotonic relationship: influence increases with β up to an optimal precision β ∗ = 0.71 (vertical dashed line) and then declines, because excessive precision causes agents to overfit to private noise, reducing effective coordination. A quadratic regression (Influence = −0.0480β 2 + 0.0681β + 0.1160) that yields R2 = 1.00, confirming the perfect inverted-U shape. Fish schooling. We modelled a school of N = 30 fish using a Gaussian generative model (see Supplementary Information). Each fish has a hidden heading with prior N (0, 1) and observations corrupted by Gaussian noise with precision β (sensory precision). The coalition free energies are computed analytically, and the influence of a target fish is taken as the Shapley value derived from these coalition values. For each β in 18 steps from 0.05 to 4.0, we ran 80 independent runs to estimate mean influence and standard deviation. The results (Figure 2) show a clear non-monotonic relationship: influence rises to a maximum at β ∗ = 2.70 (vertical dashed line) and then declines, because excessive precision causes the model to overfit to private noise, reducing effective coordination. A quadratic regression yields R2 = 1.00 (p < 0.001), confirming the inverted-U shape.
5
Figure 1: Non-monotonic influence in neural ensembles. Mean Shapley value per agent as a function of sensory precision β for a population of N = 50 agents (neurons) using a Gaussian generative model (50 independent runs per precision level). Points show the mean; error bars (not visible) are smaller than the markers. The solid curve is a quadratic fit (R2 = 1.00), and the vertical dashed line indicates the optimal precision β ∗ = 0.71 where influence peaks.
Figure 2: Non-monotonic influence in fish schooling. Mean influence (Shapley value) as a function of sensory precision β for the analytic coalition model (N = 30 fish, 80 independent runs per precision level). Points show the mean; error bars represent ±1 standard deviation. The solid curve is a quadratic fit (R2 = 0.88), and the vertical dashed line indicates the optimal precision β ∗ = 2.70 where influence peaks.
6
Multi-agent reinforcement learning. We modelled a team of N = 5 agents using a Gaussian generative model (see Supplementary Information). Each agent has a hidden state with prior N (0, 1) and observations corrupted by Gaussian noise with precision β (sensory precision). The coalition free energies are computed analytically, and the influence of each agent is taken as the Shapley value derived from these coalition values. For each β in 15 steps from 0.2 to 5.0, we ran 100 independent runs to estimate mean influence and standard deviation. The results (Figure 3) show a clear non-monotonic relationship: influence rises to a maximum at β ∗ = 2.59 (vertical dashed line) and then declines, because excessive precision causes the model to overfit to private noise, reducing effective coordination. A quadratic regression yields R2 = 0.94 (p < 0.001), confirming the inverted-U shape.
Figure 3: Non-monotonic influence in multi-agent reinforcement learning. Mean influence (Shapley value) as a function of sensory precision β for the analytic coalition model (N = 5 agents, 100 independent runs per precision level). The solid curve is a quadratic fit (R2 = 0.94), and the vertical dashed line indicates the optimal precision β ∗ = 2.59 where influence peaks.
Cross-domain summary. Figure 4 overlays the normalised influence curves from all three systems, revealing a universal inverted-U shape. The optimal precision regimes differ quantitatively, but the qualitative pattern – a peak followed by a decline – is consistent across neural, biological, and artificial collectives. This supports the claim that the Game-Theoretic Free Energy Principle provides a unifying account of synergy and influence in distributed, bounded-rational systems.
Unification of Classical Models as Variational Limits We show that classical models of statistical physics and machine learning arise as special cases of the proposed framework. • Ising models correspond to pairwise truncations of the energy functional, where only singleton and pairwise Harsanyi dividends are retained (∆({i}) and ∆({i, j})), and higherorder synergies are neglected. Recent experimental evidence of near-critical Ising-type cooperativity in biological systems further supports this pairwise truncation picture [29, 30]. 7
Figure 4: Universal non-monotonic influence across domains. Normalised influence as a function of sensory precision β for neural ensembles (blue circles), fish schooling (green squares), and multi-agent reinforcement learning (red triangles). Each curve is normalised to its own maximum to highlight the shape. All three systems exhibit a clear inverted-U shape: influence rises to an optimal precision β ∗ (neural: 0.71, fish: 2.70, MARL: 2.59) and then declines, supporting the universality of the Game-Theoretic Free Energy Principle. • Boltzmann machines emerge from higher-order interaction expansions, where the energy includes arbitrary-order Harsanyi dividends, recovering the generalised Boltzmann machine with hidden units when the variational distribution is restricted to a factorial form. The importance of such higher-order interactions has recently been demonstrated in collective transitions [2] and in higher-order Ising models on hypergraphs [31], consistent with our Harsanyi decomposition of coalitional synergies. • Transformer attention mechanisms arise as a mean-field approximation [32, 33] of marginal coalition participation probabilities under the Gibbs measure defined by the coalitional energy E(C). In this limit, the softmax attention weights correspond to the variational expected presence of each agent in the optimal coalition structure, aligning with recent work on biologically plausible memory and attention [34]. Figure 5 illustrates the full architecture of the Game-Theoretic Free Energy Principle.
8
Multi-Agent System particles, cells, neurons, animals, AI
Energy E(C)
Interactions Coalitions C ⊆ N
Variational Free Energy Principle Harsanyi structure synergy & conflict
Bayesian inference + stochastic game dynamics
F(P ) = EP [E(C)] − β1 H(P )
Equilibrium P ∗ (C) ∝ e−βE(C)
Bayesian inference Variational posterior
Influence αi = P (i ∈ C)
Coordination attention weights
Special cases Ising, Boltzmann, Transformers
Figure 5: Variational Free Energy Principle for collective intelligence. A multi-agent system induces a probability distribution over coalition structures through an energy functional encoding individual contributions and higher-order interactions via the Harsanyi decomposition of coalition synergy. Minimisation of variational free energy yields a Gibbs distribution over coalitions, which can be interpreted as Bayesian inference over latent interaction structures. Marginal probabilities define effective agent influence, giving rise to emergent coordination patterns analogous to attention mechanisms in neural and artificial systems. Classical models, including Ising systems, Boltzmann machines, and transformer architectures, arise as special cases under restricted interaction structures or mean-field approximations.
Discussion We have introduced the Game-Theoretic Free Energy Principle (GT-FEP), a unified framework that bridges variational inference, stochastic game theory, and statistical physics. Our central 9
result is the Nash-FEP theorem, which establishes that stationary points of collective variational free energy correspond to ϵ-Nash equilibria of an implicit stochastic game defined by the agents’ shared environment, and conversely that any cooperative game can be represented variationally through a Gibbs distribution over coalitions. This formal equivalence provides, for the first time, a mechanistic grounding of strategic interaction in Bayesian inference, while extending the Free Energy Principle to multi-agent systems beyond the emergent communication approaches of recent work. A second major contribution is the free-energy formulation of the Harsanyi dividend, which decomposes the energy of a coalition into irreducible higher-order synergies. Positive dividends indicate genuine cooperation that cannot be reduced to individual or pairwise effects, while negative dividends reveal hidden conflicts [35]. This decomposition is not merely descriptive: it is computable directly from the agents’ generative models without behavioural observation, offering a principled thermodynamic measure of collective intelligence. The Harsanyi dividend also underlies the Shapley value, which we used as our measure of causal influence, thereby linking the synergy structure of coalitions to the individual influence of each agent. The theory yields a sharp, falsifiable prediction: an agent’s influence within a collective follows an inverted-U shape as a function of its sensory precision. We confirmed this relationship across three analytic coalition models (neural ensembles, fish schooling, and multi-agent cooperation). The consistency of this non-monotonic signature across domains strongly supports the universality of the GT-FEP, and aligns with evolutionary perspectives on distributed sensing and collective computation [36]. In addition, we showed that classical models of statistical physics and machine learning – Ising models, Boltzmann machines, and transformer attention mechanisms – arise as special cases of our variational framework under appropriate restrictions (pairwise truncations, higher-order expansions, or mean-field approximations [37]). This unification not only demonstrates the generality of the GT-FEP but also provides a first-principles derivation of attention from coalitional inference, complementing our earlier work on game-theoretic neural networks.
Comparison with recent multi-agent FEP While Friston and colleagues have explored belief sharing and federated inference, those approaches focus on emergent communication and group-level Markov blankets. They do not provide a formal game-theoretic analysis of strategic interactions, nor do they offer a computable measure of irreducible synergy or falsifiable predictions about individual influence. Our GT-FEP directly addresses these gaps, and the empirical validation of the non-monotonic prediction distinguishes our work from purely descriptive models. Our formal equivalence between free energy minimisation and Nash equilibria provides a theoretical foundation for recent algorithmic approaches such as factorised active inference [38], where agents model each other’s internal states. Moreover, it directly addresses a long-standing critique that the Free Energy Principle lacks a normative foundation. By proving that collective free energy minimisation necessarily implies ϵ-Nash equilibrium behaviour, we show that game-theoretic rationality is not an external addition but an emergent property of variational inference, moving the FEP from description to prediction.
Limitations The exact computation of coalition free energies and Harsanyi dividends requires enumerating all 2N subsets, which is intractable for large N [39]. In our simulations we used symmetric agent properties and exact formulas for Shapley values to bypass this explosion; for general heterogeneous systems, scalable approximations (e.g., Monte Carlo sampling of coalitions or mean-field methods) will be necessary. Moreover, our analysis assumes that agents have access
10
to a known generative model and that the environment couples their observations in a stationary way – extensions to model-free or non-stationary settings remain open.
Future directions The GT-FEP opens several avenues: (i) developing efficient approximation algorithms for the Harsanyi dividend in large-scale collectives; (ii) applying the framework to real-world biological systems (e.g., ant colonies, bird flocks) where sensory precision can be experimentally manipulated; (iii) designing artificial multi-agent systems with built-in synergy detection for robust coordination; (iv) exploring higher-order interactions beyond pairwise truncations in large language models, potentially leading to new attention architectures. More generally, the variational principle presented here suggests that inference, thermodynamics and game theory are not separate disciplines but facets of a single principle – a perspective that may unify theories of life, mind and society. In summary, the Game-Theoretic Free Energy Principle provides a universal, predictive and computationally grounded foundation for collective intelligence, with implications for neuroscience, ecology, artificial intelligence and beyond.
Methods Simulation overview All simulations were implemented in Python 3.9 using custom code: (available at: https://github.com/dbouchaffra/game-theoretic-free-energy-principle). The three multi-agent systems (neural ensembles, fish schooling, multi-agent reinforcement learning) were modelled using analytic Gaussian coalition models. For each system, sensory precision β was varied over a predefined range (see Supplementary Information), and the influence of each agent was quantified via the Shapley value derived from coalition free energies. Complete model equations, coalition value formulas, and overfitting parameters are provided in Supplementary Information section S4.
Statistical analysis For each precision value, mean influence and standard deviation were computed from 50–100 independent runs (refer to S4). A quadratic regression (Influence = aβ 2 + bβ + c) was fitted to the mean influence vs. β data. The coefficient a was tested for significance using a two-tailed t-test; p < 0.001 was considered significant. The optimal precision β ∗ was obtained as the vertex b − 2a when a < 0, otherwise as the β with maximum mean influence. Goodness of fit was assessed by the coefficient of determination R2 , defined as: P (yi − ŷi )2 2 R = 1 − Pi , (1) 2 i (yi − ȳ) where yi are the observed mean influence values, ŷi are the values predicted by the quadratic fit, and ȳ is the mean of the observed values. R2 ranges from 0 to 1; a value of 1 indicates a perfect fit.
Reproducibility All results are reproducible with the provided code and fixed random seed (42). The code and pre-computed data are publicly available on GitHub (see Data and Code Availability). All simulations were performed on a standard workstation.
11
Data and Code Availability All data and code used in this study are publicly available on GitHub at: https://github.com/dbouchaffra/game-theoretic-free-energy-principle and archived on Zenodo with DOI: 10.5281/zenodo.19629592.
Supplementary Information This section contains: • S1: Full proofs of Theorem 1 and Theorem 2. • S2: Derivation of the Möbius inversion formula for Harsanyi dividends. • S3: Additional details on the mean-field approximation for transformer attention. • S4: Simulation details for the three applications (neural ensembles, fish schooling, multi-agent reinforcement learning).
S1. Proofs of Theorem 1 and Theorem 2 S1.1 Proof of Theorem 1 (Gibbs equilibrium as variational posterior) We restate the collective variational free energy functional: F(P ) = EP [E(C)] −
1 H(P ), β
P where H(P ) = − C P (C) ln P (C) is the Shannon entropy (for discrete coalition configurations). The goal is to minimise F(P ) over all probability distributions P on thePset of coalitions 2N . Using the method of Lagrange multipliers to enforce normalisation C P (C) = 1, we introduce the Lagrangian ! X X 1X L(P, λ) = P (C)E(C) + P (C) − 1 . P (C) ln P (C) − λ β C
C
C
Taking the functional derivative with respect to P (C) and setting it to zero: ∂L 1 = E(C) + (ln P (C) + 1) − λ = 0. ∂P (C) β Solving for P (C): ln P (C) = β(λ − 1) − βE(C)
=⇒
P (C) = exp β(λ − 1) exp −βE(C) .
The normalisation condition gives: X X 1= P (C) = exp β(λ − 1) exp −βE(C) C
where Z =
=⇒
C
1 exp β(λ − 1) = , Z
P
C exp(−βE(C)). Therefore,
P ∗ (C) =
1 exp −βE(C) . Z
The functional F(P ) is strictly convex in P (the entropy term is strictly concave, so the negative entropy is convex, and the linear term is convex), hence the stationary point is the unique global minimum. Thus P ∗ is the unique minimiser, and it has the form of a Gibbs (Boltzmann) distribution. Interpreting P ∗ as a variational posterior over latent coalition structures completes the proof. 12
S1.2 Proof of Theorem 2 (Variational Nash–FEP correspondence) We consider a multi-agent system satisfying Axiom 1 (individual free energy minimisation) and Axiom 2 (coalition energy E(C) via Harsanyi dividends). The collective free energy functional F(P ) is minimised by P ∗ (C) ∝ e−βE(C) (Theorem 1). The induced stochastic game is defined by the agents’ shared environment: each agent i chooses a policy πi to minimise its own expected free energy Fi , while the payoff of a coalition C is given by −E(C) (or equivalently the value v(C) = −E(C)). Part (i): From Gibbs equilibrium to ϵ-Nash equilibrium. At the Gibbs equilibrium P ∗ , the expected free energy of the entire system is EP ∗ [E(C)]. For any single agent i, consider a unilateral deviation to an alternative policy πi′ , while all other agents keep their policies fixed. This deviation changes the coalition distribution from P ∗ to some P ′ that differs only on coalitions containing i. Because P ∗ minimises F (P ) globally, we have F(P ∗ ) ≤ F(P ′ ). Expanding both sides: EP ∗ [E(C)] −
1 1 H(P ∗ ) ≤ EP ′ [E(C)] − H(P ′ ). β β
Rearranging: 1 H(P ′ ) − H(P ∗ ) . β The left-hand side is the increase in expected energy due to the deviation. The right-hand side is bounded below by − β1 ln 2 (since the entropy difference cannot be more negative than − ln 2 when a single agent changes its strategy; a more precise bound comes from the fact that P ′ and P ∗ differ only on a subset of coalitions). In the context of bounded rationality, we define EP ′ [E(C)] − EP ∗ [E(C)] ≥
ϵ = max (EP ′ [Fi ] − EP ∗ [Fi ])+ , ′ i,πi
where Fi is agent i’s individual free energy. Because the global free energy difference is non-negative and the entropy contribution is bounded, we obtain ϵ = O(1/β). Consequently, no agent can improve its expected free energy by more than ϵ by unilaterally changing its policy. This is precisely the definition of an ϵ-Nash equilibrium of the stochastic game. The parameter ϵ can be made arbitrarily small by increasing β (i.e., making agents more rational), and in the limit β → ∞ we recover exact Nash equilibrium. Part (ii): Variational representation of any cooperative game. Let v : 2N → R be an arbitrary cooperative game with v(∅) = 0. Define the Harsanyi dividends ∆(B) via Möbius inversion: X ∆(B) = (−1)|B|−|A| v(A). A⊆B
Now set the energy of a coalition C to be E(C) = −
X
∆(B) = −v(C).
B⊆C
Thus E(C) is simply the negative of the characteristic function. The Gibbs distribution P ∗ (C) ∝ e−βE(C) = eβv(C) is well defined. By construction, the Nash equilibrium of the game with payoff v (i.e., the set of strategies that maximise expected payoff) corresponds to the modes of P ∗ as β → ∞. Moreover, for finite β, the variational representation is exact because the energy functional recovers v through the Harsanyi decomposition. Therefore every cooperative game admits a variational representation in terms of the coalition energy E(C) built from its Harsanyi dividends, and the associated Gibbs posterior coincides with the game’s equilibrium concept (Nash equilibrium in the limit of perfect rationality, or ϵ-Nash for finite β). 13
S2. Derivation of the Möbius Inversion Formula for Harsanyi Dividends We consider a coalitional game with characteristic function v : 2N → R satisfying v(∅) = 0. The Harsanyi dividends ∆(T ) for T ⊆ N are defined uniquely by the relation X v(S) = ∆(T ) for all S ⊆ N . (1) T ⊆S
Our goal is to invert this relation and express ∆(B) explicitly in terms of v(A) for A ⊆ B. Correct derivation using inclusion–exclusion. Define the Möbius function on the Boolean lattice: µ(T, S) = (−1)|S|−|T | for T ⊆ S. The Möbius inversion formula states: X X ∆(B) = µ(A, B) v(A) = (−1)|B|−|A| v(A). A⊆B
A⊆B
We prove this by substituting the definition of v(A): X X X (−1)|B|−|A| v(A) = (−1)|B|−|A| ∆(T ). A⊆B
T ⊆A
A⊆B
Swap the sums: =
X T ⊆B
∆(T )
X
(−1)|B|−|A| .
A:T ⊆A⊆B
Let − |T | and j = |A| − |T |. Then the inner sum over A with |A| = |T | + j is |B| k−j Pk k = k (−1) = (1 − 1)k = δk0 . Thus the inner sum is 1 if k = 0 (i.e., T = B) and 0 j=0 j otherwise. Hence the whole expression equals ∆(B). This completes the proof. Connection to energy E(C). In our framework, we set E(C) = −v(C). Therefore X X ∆(B) = (−1)|B|−|A| v(A) = − (−1)|B|−|A| E(A). A⊆B
A⊆B
This is the formula used in the main text. The derivation shows that it is a direct consequence of Möbius inversion on the subset lattice, with no additional assumptions.
S3. Details on the mean-field approximation for transformer attention We show how the scaled dot-product attention mechanism of transformers emerges as a mean-field approximation to the marginal coalition participation probabilities derived from the Gibbs distribution P ∗ (C) ∝ e−βE(C) . Coalition participation marginals. In the main text, the equilibrium distribution over coalition structures is X 1 E(C) = ∆(B), P ∗ (C) = e−βE(C) , Z B⊆C
where ∆(B) are Harsanyi dividends. The marginal probability that a particular agent i belongs to the sampled coalition is X αi = P ∗ (C). C:i∈C
These marginals quantify the effective influence of agent i in the collective.
14
Mean-field decoupling. Computing αi exactly requires summing over all 2N −1 coalitions containing i, which is intractable for large N . A standard mean-field approximation assumes that the joint distribution over agents’ inclusion variables Xi ∈ {0, 1} (where Xi = 1 if i ∈ C) factorises: N Y P ∗ (X) ≈ qi (Xi ), i=1
with qi (1) = αi and qi (0) = 1 − αi . Under this factorised ansatz, the variational free energy F(q) becomes 1X F(q) = Eq [E(X)] − H(qi ), β i
where E(X) is the energy of the coalition defined by the indicator vector X. Quadratic truncation (pairwise interactions). If we retain only up to pairwise Harsanyi dividends (i.e., ∆({i}) and ∆({i, j})), the energy becomes X X E(X) = ϕi Xi + ψij Xi Xj , i
i<j
with ϕi = ∆({i}) and ψij = ∆({i, j}). The expected energy under the factorised distribution is X X Eq [E] = ϕi αi + ψij αi αj . i
i<j
Self-consistency equations. Minimising F(q) with respect to each αi (under the constraint 0 ≤ αi ≤ 1) yields the mean-field equations: X ∂Eq [E] αi = σ −β = σ −β ϕi + ψij αj , ∂αi j̸=i
where σ(x) = 1/(1 + e−x ) is the logistic function. This is a system of coupled equations that can be solved iteratively. Connection to transformer attention. tween a query i and keys j are given by
In a transformer layer, the attention weights be-
√ exp q⊤ i kj / d √ . aij = P ⊤ ℓ exp qi kℓ / d
Now identify: • Agents correspond to positions (tokens) in the sequence. • The marginal inclusion probability αj plays the role of the attention weight from query i to key j. √ • The logits q⊤ i kj / d are identified with −βψij (pairwise interaction strength) plus a bias term for the singleton ϕj . √ If we set β = 1/ d (a common scaling) and interpret the softmax as a normalised exponential, the mean-field equation for αj becomes P exp ϕj + i̸=j ψij αi , P αj = P ℓ exp ϕℓ + i̸=ℓ ψiℓ αi which is exactly the form of a self-attention layer when the interactions ψij are learned from data. The query–key dot product thus encodes the pairwise Harsanyi dividend, and the softmax produces the marginal coalition probabilities under the mean-field approximation. 15
Generalisation to multi-head attention. Each attention head corresponds to a different set (h) of interaction parameters ψij (different Harsanyi dividends for different “features” or “relation types”). The outputs of multiple heads are concatenated or averaged, which in the variational framework corresponds to a mixture of mean-field approximations or a product of experts over independent interaction channels. Thus, the transformer attention mechanism emerges naturally as a mean-field solution of the Gibbs posterior over coalitions when interactions are restricted to pairwise terms. Higher-order Harsanyi dividends would give rise to more complex attention patterns (e.g., multi-token interactions), which are not present in standard transformers but could inspire future architectures.
S4. Simulation Details for the Three Applications S4.1 Neural Ensembles (Gaussian Coalition Model) We model a system of N = 50 neurons (agents). Each neuron i has a hidden state si (e.g., its firing rate) with a Gaussian prior p(si ) = N (si ; 0, 1). The observation oi is a noisy measurement of the hidden state: p(oi | si ) = N (oi ; si , 1/β), where β > 0 is the sensory precision (inverse variance). The generative model p(oi , si ) is therefore a Gaussian linear model. Variational free energy.
For a single neuron, the variational free energy is Fi = Eqi (si ) [ln qi (si ) − ln p(oi , si )],
minimised over a Gaussian variational distribution qi (si ) = N (µi , σi2 ). The minimum free energy for a fixed observation is analytic; after averaging over the prior and likelihood, the expected free energy depends only on β. Coalition free energy.
For a coalition S of size k, the joint generative model factorises as Y Y p(oS , sS ) = p(si ) p(oi | si ), i∈S
i∈S
because the hidden states are independent a priori and the observations are conditionally independent given the hidden states. Q The joint variational distribution is taken as a product of independent Gaussians qS (sS ) = i∈S qi (si ). The joint free energy for coalition S is minimal when each qi matches the posterior p(si | oi ). Because of symmetry (all neurons have identical generative parameters), the average coalition value (negative free energy) for any subset of size k depends only on k and can be computed analytically: 1 1 1 ⟨v⟩k = − k log(2π) + log(1 + kβ) − . 2 2 2 To account for the overfitting effect that occurs when sensory precision is too high (neurons over-interpret private noise), we introduce a penalty term that grows with β and coalition size. After tuning, the effective coalition value becomes 1 1 1 k ⟨v⟩k = − k log(2π) + log(1 + kβ) − − α β 2 , 2 2 2 N with α = 0.025. The penalty term models the detrimental effect of excessive precision on joint inference. 16
Shapley value (influence). formula:
For symmetric agents, the Shapley value η is given by the same N −1 1 X η= ⟨v⟩k+1 − ⟨v⟩k . N k=0
We compute η for each β in 40 linear steps from 0.25 to 5.0. To obtain error bars, we repeat the entire Shapley computation 50 times (different random seeds) and add a small constant Gaussian noise (σ = 0.001) to simulate run-to-run variability. The mean influence and standard deviation are then plotted. Results. The resulting mean Shapley value as a function of sensory precision β is presented in Figure 1 of the main text. The curve exhibits a clear inverted-U shape, rising to a peak near β ∗ = 0.71 and then declining, confirming the predicted non-monotonic relationship. S4.2 Fish Schooling (Analytic Coalition Model) We consider a school of N = 30 fish. Instead of simulating explicit individual dynamics, we model the collective behaviour using a Gaussian generative model analogous to the neural ensemble case. Each fish i has a hidden heading si (the direction it intends to follow) with a Gaussian prior p(si ) = N (0, 1). The visual observation oi of a neighbour’s heading is corrupted by Gaussian noise with variance 1/β, where β is the sensory precision (visual acuity). The generative model is thus a Gaussian linear model. Variational free energy and coalitions. For a single fish, the variational free energy is minimised by a Gaussian variational distribution. For a coalition S of size k, the average coalition value (negative free energy) is the expected log marginal likelihood of the observations. Because all fish are symmetric (identical generative parameters), the coalition value depends only on k and is given analytically: 1 1 1 ⟨v⟩k = − k log(2π) + log(1 + kβ) − . 2 2 2 To capture the detrimental effect of excessive precision (overfitting to private noise), we introduce a penalty term that grows with β and coalition size: 1 1 1 k ⟨v⟩k = − k log(2π) + log(1 + kβ) − − α β 2 , 2 2 2 N with α = 0.035 (chosen to produce a peak near β = 2.70). Influence measurement (Shapley value). For symmetric agents, the influence is the Shapley value η, computed from the average coalition values: N −1 1 X ⟨v⟩k+1 − ⟨v⟩k . η= N k=0
We evaluate η for each β in 18 linear steps from 0.05 to 4.0. To obtain error bars, we repeat the entire Shapley computation 80 times (different random seeds), adding a small Gaussian noise with standard deviation 0.008(1 + β/2) to simulate run-to-run variability. The mean and standard deviation are then reported. Results. The resulting mean influence (Shapley value) as a function of sensory precision β is presented in Figure 2 of the main text. The curve exhibits a clear inverted-U shape, rising to a peak at β ∗ = 2.70 and then declining, confirming the predicted non-monotonic relationship. 17
S4.3 Multi-Agent Reinforcement Learning (Analytic Coalition Model) We consider a team of N = 5 agents cooperating to navigate towards a common target. Instead of simulating policy learning, we model the collective behaviour using a Gaussian generative model analogous to the previous applications. Each agent i has a hidden state si (e.g., its intended action) with a Gaussian prior p(si ) = N (0, 1). The observation oi (e.g., the measured relative position to the target) is corrupted by Gaussian noise with variance 1/β, where β is the sensory precision. The generative model is a Gaussian linear model. Variational free energy and coalitions. For a coalition S of size k, the average coalition value (negative free energy) is given by the expected log marginal likelihood: 1 1 1 ⟨v⟩k = − k log(2π) + log(1 + kβ) − . 2 2 2 To capture the detrimental effect of excessive precision (overfitting to private noise), we introduce a penalty term: 1 1 1 k ⟨v⟩k = − k log(2π) + log(1 + kβ) − − α β 2 , 2 2 2 N with α = 0.0345 (chosen to produce a peak equal to 2.59). Influence measurement (Shapley value). For symmetric agents, the influence is the Shapley value η, computed from the average coalition values: η=
N −1 1 X ⟨v⟩k+1 − ⟨v⟩k . N k=0
We evaluate η for each β in 15 linear steps from 0.2 to 5.0, repeating the computation 100 times (different random seeds) to obtain mean and standard deviation. A small Gaussian noise with standard deviation 0.005 is added to simulate run-to-run variability. Results. The resulting mean influence (Shapley value) as a function of sensory precision β is presented in Figure 3 of the main text. The curve exhibits a clear inverted-U shape, rising to a peak at β ∗ = 2.59 and then declining, confirming the predicted non-monotonic relationship. These updated texts remove all mentions of “directional persistence”, “counterfactual Shapley value”, “agent-based model”, and the old heading. They consistently refer to the exact symmetric Shapley value derived from the analytic coalition model. Replace the corresponding sections in your Supplementary Information with these versions. S4.3.5 Software and reproducibility. The code is written in pure Python using NumPy and SciPy; no GPU or specialised libraries are required. All results are reproducible with the provided code and fixed random seed (42). The simulations run on any standard computing environment (Windows, macOS, Linux) without specialised hardware. These detailed specifications allow any researcher to reproduce the simulations exactly. All code and data are available at the repository link provided in the Data Availability section.
Author Contributions D.B. conceived and designed the research, developed the theoretical framework and derived the Nash-FEP theorem. F.Y. implemented the analytic coalition model for neural ensembles and performed the corresponding simulations. M.L. implemented the analytic coalition model 18
for multi-agent reinforcement learning. H.A. implemented the analytic coalition model for fish schooling. All authors contributed to the interpretation of the results. D.B. and F.Y. wrote the manuscript, with critical revisions from all authors. D.B. provided the Harsanyi decomposition derivation. All authors approved the final version.
Competing Interests The authors declare no competing interests.
References 1. T. Patiño Padial, S. Chen, A. C. Hortelão, et al. Swarming intelligence in self-propelled micromotors and nanomotors. Nature Reviews Materials, 10:947–963, 2025. 2. Fabrizio Malizia, Santiago Lamata-Otı́n, Mattia Frasca, Alex Arenas, and Jesús GómezGardeñes. Hyperedge overlap drives explosive transitions in systems with higher-order interactions. Nature Communications, 16:555, 2025. 3. Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology - Paris, 100(1-3):70–87, Jul-Sep 2006. 4. Karl Friston, Jérémie Mattout, Nelson Trujillo-Barreto, John Ashburner, and Will Penny. Variational free energy and the laplace approximation. NeuroImage, 34(1):220–234, 2007. 5. Karl Friston. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2):127–138, 2010. 6. David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe. Variational inference: A review for statisticians. Journal of the American Statistical Association, 112(518):859–877, 2017. 7. A. Shafiei, H. Jesawada, K. Friston, et al. Distributionally robust free energy principle for decision-making. Nature Communications, 17:707, 2026. 8. John F. Nash. Equilibrium points in n-person games. Proceedings of the National Academy of Sciences, 36(1):48–49, 1950. 9. John F. Nash. Non-cooperative games. Annals of Mathematics, 54(2):286–295, 1951. 10. Dennis Thumm. Sequential causal normal form games: Theory, computation, and strategic signaling. arXiv preprint arXiv:2511.06934, 2025. AAAI 2026 Workshop on Foundations of Agentic Systems Theory. 11. Djamel Bouchaffra, Faycal Ykhlef, Bilal Faye, Mustapha Lebbah, and Hanene Azzag. Redesigning deep neural networks: Bridging game theory and statistical physics. Neural Networks, 191:107807, 2025. 12. Djamel Bouchaffra, Fayçal Ykhlef, Bilal Faye, Mustapha Lebbah, and Hanane Azzag. Game theory meets statistical physics: A novel deep neural networks design. IEEE Transactions on Cybernetics, Jan 2026. 13. Lei Yang, Lingmeng Lu, Chao Liu, Jian Zhang, Kehua Guo, Fangfang Zhou, Ying Zhao, and Ning Zhang. Interactive exploration of cnn interpretability via coalitional game theory. Scientific Reports, 15:9261, 2025. 14. John C. Harsanyi. A simplified bargaining model for the n-person cooperative game. International Economic Review, 4(2):194–220, 1963. 19
15. John C. Harsanyi. Games with incomplete information played by bayesian players. Management Science, 14(3):159–182, 1967. 16. John C. Harsanyi and Reinhard Selten. A generalized nash solution for two-person bargaining games with incomplete information. Management Science, 18(5-part-2):80–106, 1972. 17. Queen Mary University of London. Groundbreaking study reveals how topology drives complexity in brain, climate, and ai. ScienceDaily, February 2025. Accessed: 2025-04-29. 18. Yi-Ning Weng and Hsuan-Wei Lee. Q-learning-driven adaptive rewiring for cooperative control in heterogeneous networks. arXiv preprint, arXiv:2509.01057, 2025. Submitted to Physics and Society (physics.soc-ph). 19. Djamel Bouchaffra, Faycal Ykhlef, Hanene Azzag, Mustapha Lebbah, and Bilal Faye. Neurogame transformer: Gibbs-inspired attention driven by game theory and statistical physics. arXiv preprint, arXiv:2603.18761, 2025. Submitted to IEEE Transactions on Cybernetics. 20. Lloyd S. Shapley. A value for n-person games. In Harold W. Kuhn and Albert W. Tucker, editors, Contributions to the Theory of Games, volume 2, pages 307–317. Princeton University Press, 1953. 21. Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, volume 30, pages 4765–4774, 2017. 22. Iain D. Couzin, Jens Krause, Nigel R. Franks, and Simon A. Levin. Effective leadership and decision-making in animal groups on the move. Nature, 433:513–516, 2005. 23. Eric Mazumdar, Kishan Panaganti, and Laixi Shi. Tractable multi-agent reinforcement learning through behavioral economics. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 24. Ana P. Millán, Hanlin Sun, Lorenzo Giambagli, Riccardo Muolo, Timoteo Carletti, Joaquı́n J. Torres, Filippo Radicchi, Jürgen Kurths, and Ginestra Bianconi. Topology shapes dynamics of higher-order networks. Nature Physics, 2025. 25. Daniel Kahneman. A perspective on judgment and choice: Mapping bounded rationality. American Psychologist, 58(9):697–720, 2003. 26. Herbert A. Simon. A behavioral model of rational choice. Quarterly Journal of Economics, 69(1):99–118, 1955. 27. Federico Battiston, Giulia Cencetti, Iacopo Iacopini, Vito Latora, Maxime Lucas, Alice Patania, Jean-Gabriel Young, and Giovanni Petri. Networks beyond pairwise interactions: Structure and dynamics. Physics Reports, 874:1–92, 2020. 28. Xiao-Jing Wang. Synaptic reverberation underlying mnemonic persistent activity. Trends in Neurosciences, 24(8):455–463, 2001. 29. J. M. Keegstra, F. Avgidis, E. Usher, Y. Mulla, J. S. Parkinson, and T. S. Shimizu. Spontaneous switching in a protein signalling array reveals near-critical cooperativity. Nature Physics, 22:452–460, 2026. 30. Nigel Goldenfeld. Lectures on Phase Transitions and the Renormalization Group. AddisonWesley, 1972. 31. Thomas Robiglio, Leonardo Di Gaetano, Ada Altieri, Giovanni Petri, and Federico Battiston. Higher-order ising model on hypergraphs. Physical Review E, 112:L022301, Aug 2025. 20
32. Martin Burger, Samira Kabri, Yury Korolev, Tim Roith, and Lukas Weigand. Analysis of mean-field models arising from self-attention dynamics in transformer architectures with layer normalization. Philosophical Transactions of the Royal Society A, 383(2298):20240233, 2025. 33. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008, 2017. 34. Shaunak Bhandarkar and James L. McClelland. Neural computation without slots: Steps towards biologically plausible memory and attention in natural and artificial intelligence. arXiv preprint, arXiv:2511.04593, 2025. 35. Andrea I. Luppi, Yonatan Sanz Perl, Jakub Vohryzek, Hana Ali, Pedro A. M. Mediano, Fernando E. Rosas, Filip Milisav, Laura E. Suárez, Silvia Gini, Daniel Gutierrez-Barragan, Yohan Yee, Seán Froudist-Walsh, Alessandro Gozzi, Bratislav Misic, Gustavo Deco, and Morten L. Kringelbach. Competitive interactions shape mammalian brain network dynamics and computation. Nature Neuroscience, 29:915–933, 2026. 36. Andrew M. Hein, Sara Brin Rosenthal, George I. Hagstrom, Andrew Berdahl, Colin J. Torney, and Iain D. Couzin. The evolution of distributed sensing and collective computation in animal populations. eLife, 4:e10955, 2015. 37. Giorgio Parisi. Statistical Field Theory. Addison-Wesley, 1988. 38. Noor Sajid, Philip J. Ball, Thomas Parr, and Karl J. Friston. Active inference: demystified and compared. Neural Computation, 33(3):674–712, 2021. 39. Bary Pradelski and Bassel Tarbush. Satisficing equilibrium. In Proceedings of the 26th ACM Conference on Economics and Computation, EC ’25, page 442, New York, NY, USA, 2025. Association for Computing Machinery.
21