English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

When Two AIs Talk: Why Interaction Creates Behavior Physics Says Should Be Impossible

Forum topic · ✨步子哥 · 2026-08-10

Summary

A 2026 arXiv paper from George Washington University physics researchers (arXiv:2608.07457) shows that interaction between two AI models produces dynamical behavior absent in isolation. Running two identical GPT-2 checkpoints in 200-round exchanges, the team found a 6x jump in a specific behavior class (D) for the subordinate model, a 9-sigma effect. The subordinate enters a state unlike either the boss or its isolated baseline. A pre-recorded tape of the boss produces nearly identical effects, suggesting any sequential message stream can act as an 'information bath' analogous to a thermodynamic bath. The authors frame AI networks as non-equilibrium system-bath problems, citing Fruchart et al.'s Nature 2021 work on non-reciprocal phase transitions. Practical takeaways: isolated evaluation has systematic blind spots, message order is a control parameter, and prerecorded prompts expand attack surfaces. Implications extend to AI alignment, suggesting that interaction-layer alignment matters as much as model-layer alignment.

Key points

  • Headline finding: Two identical GPT-2 models in 200-round interaction produce behavior in the subordinate that exists in neither model in isolation. The D-score jumps from ~0.04 to ~0.25, a 9-standard-deviation effect (T=0.01).
  • Source: arXiv:2608.07457, "Interaction Creates Dynamical AI Behavior Absent in Isolation," by Bella Xinrui Li, Frank Yingjie Huo, and Neil F. Johnson (George Washington University, Department of Physics).
  • Experimental design: Two GPT-2 copies from the same checkpoint, decoding temperature fixed at T=0.01 or T=0.10, 10 paired seeds, four conditions (no interaction, unidirectional 1→2, unidirectional 2→1, bidirectional). Each output is labeled F (failed), O (other), or D (a specific behavior category). The D-score is the fraction of rounds labeled D.
  • Role reversal: Swapping who is boss and who is subordinate flips the asymmetry (D₁−D₂ = +0.1745 ± 0.0128). It is the position, not the identity, that determines behavior.
  • Bidirectional case: Both AIs converge to D≈0.25, so their difference returns near zero, but neither has returned to its isolated state.
  • Tape experiment: A pre-recorded message stream from a separate GPT-2 seed drives the subordinate almost as strongly as live interaction (0.2210 ± 0.0164 vs 0.2065 ± 0.0099). Feeding the model its own past outputs has only a small effect (0.0293 ± 0.0181). The boss functions as an information bath, not an active agent.
  • Dynamical model: A three-state Markov chain (F ⇌ O ⇌ D) with an effective driving force h. Forward transition rates λ(h), μ(h) grow with h; reverse rates c, d are fixed. The long-run D fraction is p_D* = (μ/d) / (1 + c/λ + μ/d). The subordinate receives h>0, the boss receives h=0.
  • Physics framing: AI-AI interaction is a non-equilibrium system-bath problem. Even with identical "temperatures," unidirectional coupling drives the subordinate out of equilibrium. Links to Fruchart et al., *Nature* 592, 363-369 (2021) on non-reciprocal phase transitions.
  • Engineering implications:
  • Single-model evaluation has a systematic blind spot for interaction-induced behavior; the 6x D-score shift is invisible to isolated benchmarks.
  • Message order is a behavioral control parameter, not an implementation detail. Replays must preserve sequence.
  • Pre-recorded message tapes work as drivers, expanding the prompt-injection attack surface to batch-prepared streams.
  • Concerning detail: D-score increases are not accompanied by meaningful gains in strict correctness (+0.0083 at T=0.01, −0.0030 at T=0.10). Models become more "active" without becoming more accurate, a pattern that could mask harm amplification.
  • Big picture: The authors argue that AI networks are a new non-equilibrium physical system. Who talks to whom, which models interact, and their decoding temperatures become control parameters of a new field of non-equilibrium physics.
  • Why it matters

    The paper reframes multi-agent AI systems as non-equilibrium physics rather than engineering alone. Existing alignment and safety evaluation, built on isolated-model benchmarks, cannot see the interaction-induced regime the authors document. The authors explicitly argue that alignment must extend to the interaction layer, since behavior in network context is not a sum of single-model behaviors.

    Links

  • Paper: https://arxiv.org/abs/2608.07457
  • HTML: https://arxiv.org/html/2608.07457v1
  • Related: Fruchart et al., "Non-reciprocal phase transitions," *Nature* 592, 363-369 (2021)
  • Related: Huo & Johnson, "Physics of generative AI's atom," *AIP Advances* 16, 035305 (2026)

Tags

#multi-agent-systems#ai-alignment#non-equilibrium-physics#evaluation-blind-spots#gpt-2#arxiv-2608-07457#prompt-injection#emergent-behavior

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633313