acceptodds
Withdrawn from ICLR 2027

Collective Alignment in LLM Multi-Agent Systems: Disentangling Bias from Cooperation via Statistical Physics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We investigate the emergent collective dynamics of LLM-based multi-agent systems on a 2D square lattice and present a model-agnostic statistical-physics method to disentangle social conformity from intrinsic bias, compute critical exponents, and probe the collective behavior and possible phase transitions of multi-agent systems. In our framework, each node of an lattice hosts an identical LLM agent holding a binary state (/, mapped to yes/no) and updating it by querying the model conditioned on the four nearest-neighbor states. The sampler temperature  serves as the sole control parameter. Across three open-weight models (llama3.1:8b, phi4-mini:3.8b, mistral:7b), we measure magnetization and susceptibility under a global-flip protocol designed to probe symmetry. All models display temperature-driven order-disorder crossovers and susceptibility peaks; finite-size scaling on even- lattices yields effective exponents whose values are model-dependent, close to but incompatible with the 2D Ising universality class (). Our method enables the extraction of effective -weighted couplings and fields , which serve as a measure of social conformity and intrinsic bias. In the models we analyzed, we found that collective alignment is dominated by an intrinsic bias () rather than by cooperative neighbor coupling, producing field-driven crossovers instead of genuine phase transitions. These effective parameters vary qualitatively across models, providing compact collective-behavior fingerprints for LLM agents and a quantitative diagnostic for the reliability of multi-agent consensus and collective alignment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-agent systems, large language models, statistical physics, collective behavior, AI alignment, critical behavior, opinion dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_c1kvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers an LLM agent at each site of a 2D Ising model. Each agent is asked to update the spin at its site given the states of its neighbours as a prompt. The paper then studies the statistical mechanics of this model through simulations to understand collective alignment in LLMs. The results indicate that the system does not belong to the Ising universality class.

Critical strengths

The paper reports a mostly negative result and does not overclaim Ising criticality for phi4-mini.

Critical weaknesses

I think the motivation for studying this problem is flawed. It is unclear why putting the agents on a lattice and letting the Ising dynamics run would tell us anything about their collective alignment. The fact that they do or do not update their spins according to the majority of their neighbours is highly simplified aspect of their behaviour, and the reason why this is a meaningful way of studying whether LLMs can coordinate with each other is not clarified.

If the system had behaved like a 2D Ising model, it could have been interesting, but according to the claims of the paper itself, it is clear that there is no meaningful correspondence between the Ising model and the model considered. For example, the fact that coordinated behaviour is mostly explained by the local field h~\tilde{h}, while the coupling J~\tilde{J} plays a minimal role, suggests that there is no motivation for making this correspondence. One can simply give an agent a set of opinions and check whether it agrees with the majority or not. This reveals an LLM's bias towards agreeing with a particular opinion or moving towards the majority. But putting such agents on an Ising lattice and studying the dynamics is unmotivated and does not provide any further insight whatsoever in my opinion.

I can point out several other technical issues, but since the review requires me to point out only things that may change my opinion, I will discuss only the significance of the problem. I do not think that this problem is interesting enough for this venue in the first place.

Questions for authors

Does the estimated kernel actually depend only on k, or also on the order and positions of the neighbours in the prompt?

1 comment

We thank the reviewer for the review. We agree that for two of the three models a single-agent test captures most of the behavior: on re-examination, their neighbor influence changes the answer probability by only a few percent. The lattice adds information when coupling is not negligible, as the sector asymmetry for mistral:7b suggests, and that is where we will focus future work. On your question: the prompt we used summarizes the neighbors into one of three descriptions (majority yes / tie / majority no). So, by construction, the kernel does not depend on neighbor order or position, only on this coarse-grained count. We are withdrawing the submission; please see our main comment under Reviewer Re7Q’s review.

Reviewer_pYiPOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, a collection of identical LLMs are made to give outputs -1 and 1, and their collective behavior is identified if they were behaving like an Ising model. The parameters of the Ising model are identified and studied, modeled as indicators of social conformity and intrinsic bias. The authors use three open-weight models for their study. The authors observe that these behaviors seems to be different for different LLM choices and hence characterize their identity in some way.

Critical strengths

The paper presents their setup quite clearly.

The experiments are demonstrated on three choices of LLMs.

The motivation for understanding interactions between LLMs is interesting.

Critical weaknesses

The novelty behind the idea and execution is minimal, given the venue.

Using an Ising Model, which outputs -1 and 1 to judge the bias of something like a swarm of LLMs with outputs which are generally in a much large space of outputs, seems too ambitious. One issue is the massive difference in the dimensionality of the problem. Another issue, Is completely removes the context that might arise from language.

The choice of Ising model also seems arbitrary. One could imagine many alternative models, potentially changing the message of the authors.

Questions for authors

N/A

1 comment

We thank the reviewer for the review. The reduction to a binary state is deliberate: it is the minimal setting in which bias and neighbor influence can be separated and compared against a reference model with known behavior. We agree this leaves out the context and richness of language interactions. The Ising model is used as a reference projection, not as a claim that LLM agents are spins. On reflection, the majority-vote model is the more natural reference for the prompt we used. We are withdrawing the submission (see our main comment under Reviewer Re7Q’s review). Future work will compare several reference models and richer state spaces.

Reviewer_nd9MOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work explores LLM agents that occupy a square lattice and query four neighbours to update their binary state. Simulations on finite size grids measure magnetization, susceptibility, and effective finite-size scaling exponents, with sampling temperature as the sole control parameter. A logistic fit separates intrinsic bias from neighbor influence. The three tested open-weight models exhibit different collective behavior, without convincing evidence of a common universality class or genuine 2D Ising criticality. Instead, the three models present a field-driven crossover and are dominated by the intrinsic bias rather than by the neighbor coupling.

Critical strengths

This work presents experimentally measurable metrics separating intrinsic label bias and neighbor influence. This allows to distinguish two mechanisms that can both produce population-level agreement. In particular, the work presents clear examples where the strong agreement arises predominantly from the (shared) intrinsic bias instead of the neighbor influence. The increasing use and scale of multi-agent systems make understanding their collective behaviour increasingly important. Although the proposed toy model and coarse-grained description are probably insufficient to characterise more realistic networks, the presented phenomenology (intrinsic bias vs neighbor influence) remains relevant to investigating these more realistic systems.

This work presents a clear warning about apparent criticality and naïve comparison between networks of agents and standard statistical-physics models. This warning is, in my opinion, important for future work using statistical-physics methodologies to study large ensembles of interacting LLMs.

Critical weaknesses

This work is, in my opinion, limited in several aspects, so that the quantitative conclusion have a narrow scope and their generality remains unclear. First, the choice to study a 2D grid seems primarily motivated by the comparison with the analytically solvable 2D Ising model. While a square lattice provides a useful benchmark, it is not representative of many real-world agent networks, and the solvability of the Ising model does not make the proposed LLM dynamics analytically solvable. Fully connected or sparse networks could offer both simpler analytical treatment and more realistic models. Second, the prompt can dramatically change the coupling and bias, as highlighted in the paper, but this dependence is not systematically explored in the reported results. [1,2] explore related models and report regimes where inter-agent influence plays a substantial role. My concern is not that these works are absent from the discussion (as they are preprints or recently published), but their findings (and apparently also the preliminary experiments mentioned in this manuscript) suggest that the conclusion that low-temperature alignment is "consensus by shared preference" rather than "consensus by deliberation" may be more model and prompt specific than the manuscript implies.

This work additionally presents an unexplained methodological choice. As I might be missing something, I also formulate this as a question below. There are only a few variations of the prompt: combinations of the four neighbours' binary states and the swapping of yes/no. For a fixed, memoryless protocol, the probabilities of yes/no could therefore be tabulated for each distinct prompt and temperature. Since only open-weight models are studied, these probabilities should be directly obtainable from the logits if decoding is restricted to single-token binary answers; otherwise, they could be estimated by repeated sampling of each distinct prompt. The system could then be simulated by sampling the resulting lookup table instead of repeated LLM calls, overcoming the main limitation on system size and statistical accuracy identified in the manuscript.

[1] De Marzo G, Castellano C, Garcia D. AI agents can coordinate via majority-following beyond human scale. Sci Adv. 2026 Aug 14

[2] El et al. Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents, arXiv (2026)

Questions for authors
  1. Could you clarify the global-flip mapping details ? Do you also swap the yes/no in the prompt, i.e. "reply only yes/no." --> "reply only no/yes" ?
  2. The assumption of the linear model (3) is not discussed in much detail, even though the J~\tilde J parameter is extracted from it and is interpreted as a measure of how strongly the answer depends on the neighborhood. However, there could be a nonlinear dependence on the neighbours that is not captured in J~\tilde J. Do you have evidence that J~\tilde J adequately represents the influence of neighbours, or is this an assumption in this work ?
  3. 293-294: It is claimed here that the CLT gives χ∝N/T\chi\propto N/T. But by definition 2, χ=N/TVar(m)=N/TVar(∑isi/N)=Var(∑isi)/NT=O(1)\chi=N/T Var(m)= N/T Var(\sum_i s_i /N)= Var(\sum_i s_i)/NT=O(1) if Var(∑isi)=O(N)Var(\sum_i s_i)=O(N). Furthermore Var(∣m∣)≤Var(m)Var(|m|)\leq Var(m) so that this both applies to χm\chi_m and χ∣m∣\chi_{|m|}. Could the scaling in NN instead come from the pooling across the sectors, where the law of total variance gives Var(∣m∣)=Eg(Var(∣m∣∣g))+Varg(E(∣m∣∣g))Var(|m|)= E_g (Var(|m| | g))+ Var_g(E(|m| | g)), and the second term could contribute a term of order 1 so that χ\chi can be O(N)O(N) ?
  4. Lines 357-358: could you clarify the first 2 criteria ? My understanding is that neither need to hold for Ising universality [1,2,3]. Could you distinguish these requirements from sufficient conditions associated with the standard equilibrium Ising model ?
  5. Iines 405/406: in equilibrium statistical physics, the fluctuation-response relation connects magnetization fluctuations to the response to a conjugate external field. However, it is not clear to me why a similar relation should apply here. Indeed, equilibrium is not established, the sampler temperature is not necessarily a thermodynamic temperature, etc. Could you clarify the link between the susceptibility measured in this work and the effects of small perturbations ?
  6. I reiterate here the question related to the second weakness. Why are repeated LLM calls needed when there are only a few distinct prompts, and the probabilities for each answer should be accessible as the models are open-weight ?

Other remark: The font size in the plots is quite small. Could it be increased to a size closer to that of the main text ?

[1] Grinstein, G. and Jayaprakash, C. and He, Yu, Statistical Mechanics of Probabilistic Cellular Automata, Phys. Rev. Lett. (1985)

[2] Bruce, A. D. and Wilding, N. B., Scaling fields and universality of the liquid-gas critical point, Phys. Rev. Lett. (1992)

[3] de Oliveira, M.J. Isotropic majority-vote model on a square lattice. J Stat Phys 66, 273–281 (1992)

2 comments

We thank the reviewer for an insightful review; please see our main comment under Reviewer Re7Q’s review. In brief, we agree on all the technical points. The CLT argument was wrong, and the law-of-total-variance term from sector pooling explains the N-scaling of χ. Neither a Z₂-symmetric Hamiltonian nor detailed balance is necessary for Ising universality. Out of equilibrium, χ is a fluctuation measure, not a response function. On the flip mapping: the yes/no labels were not swapped in the neighbor description, only in the recorded answer, so the protocol is not a relabeling. Your lookup-table suggestion is the right design, and our future work will use kernels tabulated from logits, with fully connected and sparse geometries. We also note the point on figure font size.

Thank you for your thoughtful response and for taking the time to revisit the analysis. I really appreciate your openness to the feedback. I still think this is an interesting and important research direction, and I look forward to seeing how the work develops.

Reviewer_Re7QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether agreement among copies of one LLM reflects mutual influence or a shared label preference. Identical yes/no agents on a periodic L×L lattice are updated asynchronously by prompting each with its four neighbours' states, with sampler temperature T as the only control parameter. The logs are analysed with a yes↔no "global-flip" protocol, a sector-pooled susceptibility χ|m|, finite-size scaling, and an effective coupling J̃ and field h̃ fitted by logistic regression on the neighbour sum k (Eq. 3). For llama3.1:8b, phi4-mini:3.8b and mistral:7b (L up to 24, 24 and 35; about 10⁹ calls), all models show crossovers with γ/ν = 1.02 ± 0.05, 1.75 ± 0.13 and 2.01 ± 0.04; the authors report h̃ ≫ J̃ and propose h̃/J̃ as a pre-deployment diagnostic.

Critical strengths

S1. Recasting "is LLM consensus cooperation or shared bias?" as a field-versus-coupling decomposition is clear and testable. The estimator (Sec. 2.4, Eq. 3) needs only update logs, is exact for heat-bath Ising dynamics and yields compact per-model trajectories (Fig. 4); together with the global-flip protocol that probes label bias (Sec. 2.2, App. A), it is a reusable black-box probe, though built from standard ingredients (logistic inverse-Ising fits, finite-size scaling).

S2. The effort is substantial and transparently costed: three models, up to seven sizes, 20 temperatures, at least six seeds and about 10⁹ calls with an explicit breakdown (App. C, p. 14), plus careful handling of sparse bins (App. B). The interpretation is cautious in places: phi4's γ/ν ≈ 1.75 is not read as Ising criticality (Sec. 3.2), and Sec. 4.1 explains why universality should not be expected.

Critical weaknesses

W1. The headline h̃ ≫ J̃ is not supported by Fig. 4 at low T: the text gives llama J̃ ≈ 0.35 and h̃ ≈ 2.8 "at the lowest temperature" and phi4 "J̃ ≈ 0 throughout" (p. 6), but at T = 0.05 Fig. 4 shows llama J̃ ≈ 0.9 (above 0.4407) with h̃ ≈ 2.35 and phi4 J̃ ≈ 0.33; the quoted values match T = 0.1. With |k| ≤ 4, h̃ should be compared with ≈ 4J̃, which exceeds h̃ at T = 0.05 and is only ≈ 2× smaller at T = 0.1, contradicting "nearly an order of magnitude" and the h̃/J̃ ≲ 1 threshold (Sec. 4.5). The fits also conflict with the dynamics: mistral's T = 1.2 values imply P(s′ = +1) ≲ 0.73 with aligned neighbours, yet its g = +1 sector locks at m ≈ 1 (Fig. 5). Suggestion: correct the text, report per-sector (J̃, h̃) with CIs, and check the fits by simulating the fitted dynamics.

W2. The exponent interpretation needs revision. Attributing mistral's γ/ν ≈ 2 to the central limit theorem (Sec. 3.2) is incorrect: with Eq. 2, CLT fluctuations give χ = O(1), the γ/ν = 0 case (p. 5); likewise phi4's χmax should not grow with L if J̃ ≈ 0, yet it grows about five-fold (Fig. 3b). Pooling both flip sectors (Sec. 2.3) adds a term ∝ N whenever ⟨|m|⟩ differs between sectors; mistral's sectors at m ≈ +1 and −0.5 (Fig. 5) give χ ≈ 0.05–0.06N, matching χ ≈ 80 at L = 35. Sizes are narrow (L ≤ 24 for llama and phi4, three sizes for mistral), burn-in (1.4L² sweeps) is short against relaxation of order L², and odd L is excluded on an antiferromagnet argument (App. C). Suggestion: compute per-sector χ, report all sizes, and validate the pipeline on a synthetic kinetic Ising model.

W3. Controls and ablations are missing: there is no null model (no-neighbour or shuffled-neighbour prompt; voter baselines are deferred to Sec. 5), and no ablation of prompt wording (which "can dramatically alter the (J̃, h̃) balance", p. 9), label tokens, neighbour order or sampler settings. The prompt itself is underspecified (neighbour list elided, p. 3; "+/−" in Fig. 1 versus yes/no states), and invalid-reply handling, decoding parameters and quantization are unreported; under top-k/top-p truncation, "T → ∞ gives uniform sampling" (p. 8) does not hold. Suggestion: add these controls, at least two prompts and label pairs, and invalid-reply rates.

W4. One content-free yes/no question, one prompt and three 3.8–8B models, with no ground-truth task, cannot support the conclusion that the collective "provides no additional epistemic value over a single query" (p. 8), nor its extension to debate, self-consistency and voting (pp. 2, 9). Suggestion: show on a task with known answers that h̃/J̃ predicts consensus accuracy, add larger models, or narrow the claims.

Questions for authors

Q1. What are γ/ν and χ(T) when χ is computed within each flip sector from within-run fluctuations, and how does the sector gap in ⟨|m|⟩ scale with L?

Q2. How are the two sectors combined in the logistic fit (pooled, relabelled or separate)? Please confirm the T = 0.05 values in Fig. 4 and report per-sector (J̃, h̃).

Q3. The prompt omits the agent's own state, so if the log-odds are linear in k the dynamics is exactly heat-bath Ising and obeys detailed balance. Can you report per-k empirical log-odds against Eq. 3 and test dependence on neighbour arrangement at fixed k? This would directly test the non-equilibrium claim (Sec. 4.1, App. C) against Song et al. (2025), who report detailed balance in LLM-driven agents.

Q4. How do (J̃, h̃) and γ/ν change under other wordings, label pairs (e.g., A/B), neighbour orderings and top-k/top-p settings? What fraction of replies is unparseable at high T, and how are such replies handled?

Q5. How does the diagnostic relate to conformity studies of multi-agent LLMs (Weng et al., ICLR 2025) and to emergent collective bias in LLM populations (Ashery et al., Sci. Adv. 2025)? Does a high h̃/J̃ predict lower consensus accuracy on a task with ground truth?

1 comment

We thank the reviewer for a careful and constructive review. Several of the concerns raised are correct, and a re-examination of our code prompted by this review confirmed them.

(i) The central-limit argument for γ/ν ≈ 2 was wrong. Pooling the two flip sectors adds a term proportional to N to χ_|m| whenever ⟨|m|⟩ differs between sectors, and this term accounts for the growth of the susceptibility peaks in all three models. The reported exponents are therefore not critical exponents.

(ii) The J̃ values quoted “at the lowest temperature” were the T = 0.1 values, misread from Fig. 4. At T ≤ 0.1 almost all outcomes are “yes”, so the field and coupling cannot be separated by the fit. We agree that h̃ should be compared with 4J̃. The “order of magnitude” statement and the h̃/J̃ threshold do not hold.

(iii) On Q2: the fits in Fig. 4 used only g = +1 runs. The Mistral fit is inconsistent with its m ≈ 1 sector, as you note, and is not a valid estimate.

(iv) On Q3 and the flip protocol: as implemented, the global flip inverted the model’s answer but did not relabel the neighbor description shown to the model. It is therefore not a Z₂ relabeling. The asymmetry between sectors is instead a direct test of neighbor influence. Also, the prompt actually used summarizes the neighbors into three descriptions (majority yes / tie / majority no), so the kernel cannot be linear in k, and Eq. 3 is only a two-parameter projection of a three-level kernel.

(v) On Q4: invalid replies were not logged, so we cannot report their rate. We agree that controls, ablations over prompts, labels and sampler settings, and larger models are needed. We also agree that the claims about debate, self-consistency and voting go beyond what a content-free task can support. We will engage with Weng et al. and Ashery et al. in future work.

We are withdrawing the submission. A corrected account will be made available separately. Future work will tabulate the kernel directly from logits, as also suggested by Reviewer nd9M, and add the controls and a task with ground truth.

Comments on the paper

1 comment

We thank the four reviewers for their careful reading. Several of the concerns raised are correct, and a re-examination of our code prompted by these reviews confirmed them. In particular: (i) the central-limit argument for γ/ν ≈ 2 was wrong, and pooling the two flip sectors adds a term proportional to N to χ_|m|, which accounts for the growth of the susceptibility peaks; the reported exponents are therefore not critical exponents. (ii) The J̃ values quoted at the lowest temperature were misread from Fig. 4, and the low-temperature fits are not identifiable. (iii) In answer to Reviewer nd9M’s question, the global flip as implemented inverted the model’s answer without relabeling the neighbor description, so it is not a Z₂ relabeling; the sector asymmetry is instead a direct test of neighbor influence. (iv) The even-L justification and the stated conditions for Ising universality were incorrect. We are withdrawing the submission. A corrected account will be made available separately, and we plan to follow the reviewers’ suggestions (tabulating the kernel from logits, controls and ablations, larger models, and a task with ground truth) in future work.

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

Related papers

Discussion (0)

Sign in to comment.