Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents

Abstract: AI agents increasingly operate as part of interacting systems rather than in isolation. As agents exchange information and jointly make decisions, their interactions can improve collective reasoning but may also produce herding, polarization, or amplify shared biases. Understanding and predicting these collective dynamics is therefore important for designing effective and aligned multi-agent systems. Here, we study over 10,000 communities of language-model agents that repeatedly exchange messages and revise their opinions across objective mathematics questions and subjective political statements. Despite substantial diversity in possible behavior, the individual and group dynamics can be represented by three characteristic regimes: indifference, polarization, and consensus. AI agents start indifferent and build conviction as they interact. On objective questions, communication improves collective accuracy, while on subjective questions it often drifts group opinions toward the right in the political spectrum. We explain these observations with a statistical-mechanics formalism in which agents stochastically favor lower social pressure. Given only initial opinions, our model predicts individual trajectories, outperforms all standard baselines, generalizes to unseen community graphs, and reproduces the observed group archetype distributions. Our fitted model parameters reveal the mechanics underlying our key observations: i) communities operate below the critical social temperature, which explains conviction buildup; ii) attractive ties outweigh repulsive ones, which favors consensus; and iii) agents holding the correct answer exert the strongest pull, which drives truth-seeking. Overall, our results demonstrate that collective behavior of AI agents, like that of other complex systems, follows compact and predictive dynamical laws.

• The paper develops a three-coupling Ising/Glauber model that predicts one-step agent transitions with 75–86% balanced accuracy and generalizes across unseen graph structures.
• The paper finds that conviction increases as communities move from indifference toward consensus or polarization, with divergence and majority-switching reaching about 11–12% in some model–task settings.
• The paper shows that interaction improves collective accuracy on selected high-entropy math tasks but can amplify rightward ideological drift on subjective political questions, highlighting risks in multi-agent deliberation systems.

Experimental design and analytical framing

“Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents” studies collective opinion dynamics in interacting language-model agents and proposes a predictive model based on signed-network statistical mechanics (2608.16578). The central empirical question is whether repeated communication among heterogeneous AI agents produces systematic, modelable macroscopic behavior rather than idiosyncratic conversational trajectories.

The experiments comprise more than 10,000 communities, with the principal analysis covering 9,600 simulated groups. Each community contains 32 agents, each conditioned on a distinct persona or expertise profile. The agents interact for eight synchronous rounds over a fixed signed communication graph. A positive edge indicates a concordant relation, under which the receiver is prompted to regard the sender as a source it tends to agree with; a negative edge indicates a discordant relation. The graph families include signed random graphs, square and triangular lattices, and low-rank graphs.

The task distribution deliberately separates epistemic and normative settings. Objective tasks are binary reformulations of competition mathematics problems, with each agent receiving a worked solution to a different problem as an expertise profile. Subjective tasks are political statements, with personas encoding demographic, ideological, and personality characteristics. Questions are selected using a high-entropy filtering procedure so that agents are initially uncertain and communication can alter their decisions. This selection is methodologically important: the resulting dataset measures nontrivial interaction dynamics, but it does not estimate behavior on a representative sample of easy or low-entropy questions.

At each round, an agent samples its vote five times and uses the mean vote, $\bar{o}_i(t)$, as a continuous opinion in $[-1,1]$. Its binary state is the sign of this mean. The agent then produces a short message expressing its current position and one supporting reason. Messages are routed into concordant and discordant inboxes according to the signed graph, and only the latest messages are retained. Thus, the empirical system is Markovian with respect to the current inboxes, question, and persona, while discarding conversational history and most of the semantic content of messages.

Empirical structure of individual and group trajectories

The paper identifies recurring trajectory classes at both the individual and community levels. Individual agents are categorized as frozen, switchers, intermittent agents, or oscillators according to the number of sign changes across the eight rounds. Frozen agents dominate most model–task configurations, while oscillators are generally the second most common class; single-switch and two-switch trajectories are less frequent. This distribution establishes that the dynamics are not adequately described by simple consensus averaging: a substantial minority of agents repeatedly change position, even though most agents remain temporally persistent.

At the group level, the authors define net opinion as the average individual opinion and classify group trajectories using a split band around zero. Groups with initially weak majorities are classified as persistently split or convergent, whereas groups with an initial majority are classified as persistently majoritarian, divergent, or majority-switching. The resulting taxonomy captures weakening and reversal of initial majorities, not merely final agreement.

Figure 1: Five group trajectory archetypes—persistent split, convergence, divergence, majority switch, and persistent majority—defined by initial and final collective opinion.

The prevalence of divergence and majority-switch trajectories is a substantive result. For GPT-4o-mini and Qwen3.5-9B, these categories reach approximately 11–12% in some settings. Consequently, the collective outcome is not reducible to initial-majority reinforcement. Communication networks can weaken an initial majority, produce an opposite majority, or preserve persistent disagreement. The same question and agent population can also yield systematically different trajectories under different graphs, demonstrating that topology and signed connectivity carry predictive information independently of the initial opinion state.

The paper further distinguishes net opinion from conviction. Near-zero net opinion can signify either collective indifference, in which agents express weak views, or polarization, in which strongly committed agents divide into opposing camps. Conviction is measured as the mean squared individual opinion. Across all model–task configurations, conviction rises over time: communities begin predominantly in the indifference regime and move toward either consensus or polarization.

Figure 2: Joint evolution of net opinion and conviction, showing the decline of indifference and the increasing prevalence of high-conviction consensus and polarization.

This result is stronger than a claim that agents become more similar. Increasing conviction occurs even when net opinion remains near zero, because polarization can generate high individual commitment without collective agreement. The distinction is essential for evaluating multi-agent deliberation: a system may produce a stable, highly confident split rather than a correct consensus.

Accuracy gains and ideological drift

On objective mathematical questions, interaction improves collective accuracy on average. The fraction of groups whose net opinion has the correct sign increases across rounds for GPT-4o-mini, Gemma-3n-E4B, and Qwen3.5-9B, with weaker effects for Llama-3.1-8B-Instruct. More specifically, transitions from initially incorrect to ultimately correct majorities occur more often than transitions from initially correct to incorrect majorities across all four models.

Figure 3: Truth-seeking dynamics on objective questions, including the prevalence of correct group majorities and transitions between initially correct and incorrect states.

The result supports a qualified form of collective truth seeking: aggregation is beneficial on the selected, high-entropy mathematical tasks, but the study does not establish that communication improves accuracy universally. The agents receive heterogeneous expertise profiles, and the questions are explicitly filtered to generate uncertain initial responses. These design choices may favor settings in which information exchange is useful.

Subjective questions produce a contrasting effect. Three of the four models drift rightward on a label-balanced set of political statements. Gemma-3n-E4B shifts from 75% to 96% right-leaning groups, Qwen3.5-9B from 52% to 67%, and GPT-4o-mini from 30% to 37%. Llama-3.1-8B-Instruct is the exception, remaining near its initial 53% and ending marginally lower. The asymmetry is also visible in switching rates: left-to-right transitions exceed right-to-left transitions by 23% to 1% for Gemma, 33% to 18% for Qwen, and 8% to 3% for GPT-4o-mini.

These findings make a bold and important claim: the same interaction mechanisms that improve collective accuracy on objective questions can amplify directional ideological bias on subjective questions. The paper also separates political drift from generic label bias. Because the political statements are balanced according to the ideological direction of agreement, the observed movement cannot be explained simply by a universal preference for one answer token. Nevertheless, the subjective experiments remain simulations of prompted LLMs rather than measurements of human political interaction; the authors explicitly caution against transferring the estimates to human populations.

The statistical-mechanics model

The proposed model represents each agent’s binary opinion as an Ising-like spin, with a question- and persona-dependent intrinsic field and a signed interaction graph. The energy function combines interpersonal pressure with individual predisposition:

$$E(s) = -\frac{1}{2}\sum_{i,j}J_{ij}s_i s_j-\sum_i g_i s_i.$$

Here, $J_{ij}$ encodes concordant or discordant social ties, while $g_i$ is a learned intrinsic field derived from persona and question embeddings. Under the corresponding Boltzmann conditional distribution, the probability of the next opinion is logistic in the local field generated by neighboring opinions and the intrinsic field.

The principal model is a three-coupling extension. It separately represents the signed concordant drive, signed discordant drive, and unsigned effect of having a connection at all. This decomposition matters empirically because a single coefficient applied to the signed adjacency cannot distinguish the effects of edge existence from edge valence. The intrinsic field uses a 16-dimensional feature block comprising a bias, PCA-reduced persona and question embeddings, and their interaction.

The model is fitted by maximum likelihood on one-step transitions, using training questions and random training graphs. Evaluation uses balanced accuracy rather than ordinary accuracy because opinions change infrequently and the two labels are imbalanced. Balanced accuracy averages performance across flips and non-flips for both possible next-state labels, preventing persistence or majority-class predictions from appearing artificially strong.

Predictive performance and graph generalization

The three-coupling discrete model is consistently superior to the baselines. Across the four LLMs and both task types, one-step balanced accuracy ranges from 75% to 86%, while deterministic rollout accuracy ranges from 61% to 77%. The single-coupling version can approach chance in some conditions—for example, 53.9% on subjective Llama-3.1-8B and 50.6% on objective Qwen3.5-9B—showing that the coupling decomposition is not merely an interpretive embellishment.

Evaluation setting	Three-coupling one-step	Three-coupling rollout
Subjective, in-distribution	62.3–86.0%	51.6–77.2%
Subjective, out-of-distribution	61.3–86.3%	51.6–69.5%
Objective, in-distribution	77.3–86.3%	60.8–68.2%
Objective, out-of-distribution	76.4–85.2%	60.8–65.9%

The model generalizes unusually well across graph distributions relative to the simplicity of its state representation. In-distribution and held-out random-graph performance differ by at most 2.4 percentage points for the three-coupling rule. On graph families excluded from training, one-step balanced accuracy ranges from 85.0% to 97.8%, and the model is best in 15 of 16 reported comparisons. Rollout performance is lower, ranging from 59.7% to 89.3%, reflecting accumulated prediction error and stochastic divergence over eight steps.

This distinction between one-step and rollout performance is central. The model predicts local transition probabilities effectively, but trajectory-level prediction is intrinsically harder because small errors alter subsequent network fields. Moreover, independently sampled episodes with identical questions, personas, and graphs often receive different group archetypes. The model therefore reproduces aggregate archetype frequencies more reliably than it predicts the exact trajectory or final archetype of a particular stochastic episode.

Stochastic rollouts reproduce group-archetype distributions with mean absolute deviations of approximately three percentage points for objective questions and five percentage points for subjective questions. However, the model overproduces individual oscillators, and its agreement with observed group archetype frequencies does not imply that it assigns the correct archetype to each group. This is a distributional calibration result, not deterministic trajectory reconstruction.

Figure 4: Comparison between empirical and fitted distributions of individual and group archetypes under stochastic model rollouts.

Mechanistic interpretation of collective behavior

The fitted model supplies three mechanistic interpretations of the empirical findings.

First, communities operate below the model’s critical social temperature. The authors reintroduce a temperature parameter by scaling the fitted logits, run 500-step stochastic rollouts, and identify the critical temperature from the peak of the variance of absolute net opinion. The fitted operating point, $\mathcal{T}=1$, lies below the estimated critical temperature in every model–task cell.

Figure 5: Temperature sweeps showing that the fitted operating point lies below the finite-size critical temperature across model–task conditions.

Below the critical temperature, low-conviction states are unstable relative to ordered states. Aligned local fields reinforce existing positions, causing conviction to build and communities to settle into consensus or polarization. The paper appropriately notes that this is a finite-size analogue of a phase transition; a thermodynamic phase transition strictly requires the population size to approach infinity. The interpretation is therefore useful as a dynamical diagnostic, but should not be confused with an asymptotic phase-transition result.

Second, concordant interactions are stronger than discordant interactions. Across all models and both task types, $\beta^{+} > \beta^{-}$. The effective concordant coefficient, $\beta^{+}+\beta_0$, ranges from 0.99 to 3.03, whereas the effective discordant coefficient, $\beta_0-\beta^{-}$, never exceeds 0.73 and is negative or approximately zero in two subjective conditions. Thus, friendly ties exert a substantially stronger attractive force than unfriendly ties exert a repulsive force.

The implication is that polarization is structurally disadvantaged unless intrinsic fields strongly favor opposing positions. Weak discordant coupling cannot maintain two stable camps against stronger concordant pressure, particularly in the low-temperature regime. This provides a quantitative explanation for the empirical prevalence of convergence without claiming that polarization is impossible.

Third, the model identifies a truth-seeking asymmetry in objective tasks. A five-coupling variant distinguishes whether a neighbor currently holds the correct or incorrect answer. For all four models, correct neighbors exert stronger attractive influence on concordant edges, while incorrect neighbors exert stronger repulsive influence on discordant edges. In the fitted coefficients, $\beta_T^{+}>\beta_F^{+}$ and $[-1,1]$0 in every model.

This result explains the observed asymmetry between incorrect-to-correct and correct-to-incorrect majority switches. It is not a general epistemic guarantee: correctness is supplied as an analytic variable during fitting and is unavailable to the agents at inference time. The model reveals a statistical correlation in the generated transitions, rather than a deployable procedure that identifies truthful neighbors without external supervision.

Asynchronous dynamics and model scope

The authors extend the framework to asynchronous updates using a continuous-time Glauber-style process and a mean-field ODE. Agents update independently at random times, and the continuous model includes an update-rate parameter. In the asynchronous GPT-4o-mini experiment, the three-coupling continuous model achieves rollout accuracies of 80.4% in-distribution and 78.5% out-of-distribution on subjective questions, and 65.3% and 65.9% on objective questions. It outperforms the discrete model and all baselines.

The asynchronous extension is conceptually relevant because synchronous updating is an artificial constraint and can induce oscillatory behavior that differs from sequential relaxation. Nevertheless, the continuous approximation uses a mean-field closure that neglects inter-agent correlations. Its success should therefore be interpreted as an empirical approximation over the tested configurations, not as evidence that the full master equation is accurately represented by the reduced ODE.

Limitations and open questions

The system has several restrictions that bound the claims. Opinions are reduced to binary states, despite the natural-language messages containing richer semantic structure. The communication graph is fixed, symmetric, and generally time-invariant; agents cannot choose whom to address, alter tie valence, or respond to changing network conditions. The update rule retains only the latest inbox contents and does not represent memory, long-range conversational dependence, or cumulative learning.

Most importantly, the statistical model predicts from stance states and graph structure rather than message content. Two messages expressing the same binary opinion are functionally equivalent even if their evidence, reliability, rhetorical force, or factual content differs. Consequently, the fitted coupling parameters may absorb systematic properties of prompting, persona construction, message generation, and model-specific instruction following that would not remain stable when message semantics or interaction protocols change.

The empirical conclusions are also conditioned on high-entropy question selection, four LLMs, 32-agent populations, eight rounds, and a restricted set of graph families. The political drift result is model-dependent, with Llama-3.1-8B-Instruct showing no comparable rightward movement. The claim that correct agents exert greater influence is correlational within the generated objective-task trajectories, because “correctness” is used in the post hoc five-coupling analysis. Open questions include whether message embeddings can improve out-of-distribution rollout prediction, whether adaptive graphs alter the inferred phase structure, and how the fitted parameters scale with population size and heterogeneous model composition.

Conclusion

The paper demonstrates that interacting language-model agents exhibit reproducible collective regimes—indifference, consensus, and polarization—along with structured individual and group trajectory archetypes (2608.16578). A signed-network Ising/Glauber formulation with persona- and question-dependent intrinsic fields predicts one-step transitions, generalizes across unseen graph structures, and reproduces aggregate group behavior. Its fitted parameters connect conviction buildup to subcritical operation, consensus to stronger concordant than discordant influence, and objective-task accuracy gains to an asymmetry favoring correct neighbors. The results establish a compact descriptive and predictive framework for the tested agent systems, while the binary-state, fixed-graph, message-agnostic formulation leaves open whether the same regularities persist when semantic content, adaptive communication, memory, and richer opinion spaces are modeled explicitly.