acceptodds
Under review as a conference paper at ICLR 2027

Emergent Equilibrium: A Measurable Theory of Strategic Convergence in LLM Multi-Agent Dialogue, with a Closed-Form Critical-Temperature Bound

Abstract

When multiple LLM agents debate, negotiate, or deliberate, do their dialogues converge—and when do they fail to? To this end, we give a formal answer and derive a closed-form condition for the transition. We formalize multi-agent dialogue as a discrete dynamical system on a semantic belief manifold, construct an explicit dialogue operator from cross-attention layers, and introduce the Ideal Group Introspection (IGI) framework—a two-parameter learnable model whose residual error yields the first measurable criterion for strategic convergence wherever those residuals are identifiable. Within this framework, we prove that an approximate emergent equilibrium—satisfying Nash strategic stability, geometric dialogue convergence and consensus coupling—exists under mild regularity conditions, via a sandwich argument linking Glicksberg and Brouwer, the profile delivered being -close to the stage game's quantal response equilibrium, and Nash as ; we establish spectral-radius convergence conditions and an attention contraction lemma, and show that the dialogue undergoes a phase transition whose critical sampling temperature is bounded above by the closed form , where ( the value and key projection norms, the agents, the key dimension) is the architectural constant that the reduced-operator limit recovers exactly. Crucially, this makes a one-sided bound on instability onset computable from stored weights under one decomposition assumption, and predicts critical slowing near —a prediction our sweeps do not confirm, a null result we report with its power. Validating across 15 models (four families, 0.6B to trillion-parameter scale) in strategic and tax-compliance games, we uncover three regularities that our theory explains. First, in the compliance framing frontier closed-source models exhibit sycophantic cooperation—complying even when defection is strictly dominant, while defecting in the abstract game—and so fall outside the IGI regime there; because this deviation is systematic rather than random, it defines an interpretable four-way taxonomy of model behaviour instead of a failure of the theory. Second, strategic behaviour is protocol-dependent rather than a fixed model property: a single checkpoint spans both endpoints, from complete cooperation in the tax framing to zero against randomised opponents. Third, cooperation varies non-monotonically with model scale within the mid-size regime, where strategic diversity is highest. Taken together, our results provide a rigorous, empirically testable foundation for predicting when LLM multi-agent deliberation is strategically meaningful—and when it is not.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM multi-agent dialogue, strategic convergence, dynamical systems, spectral radius, phase transition, critical temperature, emergent equilibrium, agent cooperation, tax compliance

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xVXVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper attempts to build a theoretical foundation for multi-agent LLM dialogues to answer when and how such LLM dialogues converge. The authors derive a boundary computable from model weights as a sufficient condition under which LLM dialogues stabilize.

Critical strengths

S1. The research question under exploration is interesting and important.

S2. The authors derive a boundary computable from model weights as a sufficient condition under which LLM dialogues stabilize.

S3. The empirical analysis covers 15 models, which is many.

Critical weaknesses

W1. The theoretical analysis appears sound but unreadable. The authors just put the maths there without clear explanations. Also, the practical significance of the results has not been clearly established.

W2. The paper appears to show strong signs of vibe writing, with numerous corrections to its experimental methodology, empirical results, and theoretical analysis, making it extremely difficult to understand what the paper is actually trying to say.

W3. The computable boundary overstates the local sensitivity by about 7×1037\times 10^{3} and thus is overly loose. Only 2 of the 22 checkpoints fit the boundary.

W4. The key assumption that the agents' beliefs are updated via cross-attention lacks justification and evidence.

Questions for authors

Q1. Why does this paper involve a lot of corrections?

Q2. What justifies the assumption that belief updates in multi-agent LLM dialogue can be modeled using cross-attention?

1 comment

I cannot understand what either the paper or the rebuttal is trying to say. I believe I am talking to an out-of-control AI rather than a responsible human being. I therefore decline to continue this discussion and will maintain my original score.

Reviewer_NLMQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether LLM agents can converge on stable, rational decisions through discussion. The authors develop a theoretical framework for convergence and test it using the Prisoner's Dilemma and a tax-reporting game. They find that the agents' decisions depend on framing and communication protocols, while larger LLMs do not necessarily perform better. Although the theory predicts a phase transition and critical slowing down, neither is observed in their experiments.

Critical strengths

The research question is important, and linking convergence in dialogue to model properties is a promising direction. The authors are also open about the paper's issues, including reported defects, a misattributed cooperation rate, and the failure to observe critical slowing down.

Critical weaknesses
  • The theory and experiments are not well connected. 'The closed-form critical-temperature bound' in the title is not actually derived from the stated assumptions. In Appendix A.3 (Step 6), the authors describe the argument leading to σVσKN/dk\sigma_V\sigma_K N/\sqrt{d_k} as 'a working simplification rather than a derivation,' and A.5 finds it loose by roughly 7×1037\times10^3. A phase transition and critical slowing near TcT_c are the two predictions one can actually check, and neither appears in the swept range (D.2, D.3). Under the corrected action rule, εIGI\varepsilon_{\mathrm{IGI}} is identified at only one of the 22 checkpoints (Table 7). Proposition C.1 makes the gap sharper: for the payoffs used here, the residual is a monotonic function of the cooperation rate, so it is not an independent criterion. The authors note this themselves.

  • The existence argument in Theorem 5.2 is incomplete. Appendix A.1 says that dMd_M is not a metric and that the triangle inequality is not used, but Eq. (12) is a triangle inequality in dMNd_{M^N}, and both the contraction (9) and the Cauchy argument are stated in dMd_M.

  • The paper has several presentation issues. Table 14 is empty, and Appendix D states that the abstract was written before the experiments. The same points about identifiability and estimator power are repeated across C.2, C.6, D.1, D.3, and D.9. Several figures are also too small, making their labels difficult to read.

Questions for authors

I would appreciate clarification if I have misunderstood any of these points.

Reviewer_wnaHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether AI agents that repeatedly talk and make decisions together eventually settle into a consistent pattern, and whether those decisions benefit each agent. It proposes a mathematical explanation and tests how agents behave under different instructions and amounts of randomness. The experiments show that instructions and interaction setup strongly affect behavior, but they do not confirm the paper’s main predictions. Some central mathematical arguments also have gaps, so the proposed explanation is not yet convincing

Critical strengths

The paper asks an important question: when AI agents keep interacting, do they settle on sensible decisions or simply reinforce one another’s mistakes? Its experiments show that the same AI can behave very differently depending on how the situation is described and how the interaction works, and the authors openly report when their predictions fail.

Critical weaknesses

The paper’s main mathematical claims have gaps, so it does not convincingly establish when interacting AI agents will settle into stable behavior:

  • In appendix A.1, the authors acknowledge that their measure of how different two beliefs are, lacks the triangle inequality. But Eq. 12 uses the inequality to prove agents converge.
  • The temperature formula in Appendix A.3 drops the size of the agents' internal representations.

Also, the experiments fail to confirm the predicted turning point, and the connection between its simplified model and how actual AI agents behave are unclear.

Finally, I think the paper isn't very well written. The graphs are hard to read, the text is very dense and indicative of AI-writing, and very jargon-heavy. The manuscript would benefit from simplification.

Questions for authors

Your formula gives an upper limit on instability. What establishes that agents actually approach that limit and slow down near the predicted temperature, especially since the experiments did not observe this?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.