{"id":"304ccc63-f5c0-4753-a015-6b43ede5ff61","arxiv_id":"2608.02827","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Collective bias in multi-agent LLM debates emerges as a finite-N rounded mean-field phase transition controlled by the ratio of conformity to sampling temperature.","lead":"Multi-agent debates between LLMs can lock into biased consensus at low sampling temperatures, even when each model's initial bias is small. A spin-model-based analysis predicts this as a finite-population version of a phase transition, with temperature, conformity, and network sparsity as control knobs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LLM experiments never vary N, so the finite-N rounding that anchors the phase-transition claim is unmeasured; the observed T-crossover could be single-agent stochasticity.","rationale":"The reader's weakest_assumption is that LLM choices follow the Gibbs update in Eq. (5). I agree that this mapping is not independently validated, but the more immediately load-bearing gap is empirical: the paper's distinctive quantitative prediction is the 1/sqrt(N) rounding and the lambda rho N/T threshold, and no LLM experiment varies N. The T-b phase diagrams in Figs. 2 and 10 are consistent with the theory, but they are also consistent with a purely single-agent temperature effect: at low T, any stochastic LLM with a logit bias chooses the biased token with near-unit probability, so a population of non-interacting identical agents would also show |m|~1. The N-dependence in the paper's own Fig. 9 comes from simulations of the model, not from LLMs, so it does not validate the mapping. A clean N sweep with held-out prediction and a no-debate control would settle whether the finite-N phase-transition framing is doing real work. This is an addressable experimental gap, so it does not by itself overturn the paper; rather, it reinforces the conditional recommendation: the phase-transition claim should not be stated as established until N-scaling is tested. I credit the paper with reproducible code, a standard derivation, and careful controlled experiments on design choices; the concern is about what the current data can distinguish, not about the internal derivation.","tokens_in":27518,"tokens_out":6715,"duration_ms":68711,"concrete_test":"Run the Binary Choice task with one fixed LLM (e.g., GPT-4.1 Nano) and fixed logit bias b at temperatures around the observed crossover, for N=3, 6, 11, and 20 agents with the existing Du et al. protocol, using at least 20 seeds per condition. Fit (lambda, h) on N=11 data and do not refit. Check (i) whether the measured crossover location and width in T change with N as predicted by Eq. (5)'s 1/sqrt(rho N) noise, and (ii) whether the predicted |m(R)| at N=20 from the N=11 fit matches. Include a no-debate control (agents never shown others' answers) to confirm consensus exceeds the non-interacting baseline. If no N-sharpening is observed, the finite-N phase-transition central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim rests on the finite-size version of a phase transition: Eq. (5) has noise eta(t)=O(1/sqrt(rho N)), so the crossover should sharpen and the threshold should shift toward lambda rho N/T ~ 1 as N grows. Section 6.1 and Appendix C.2, however, report synthetic LLM debates only at N=11, and the real-world tasks use N=6 and N=10. Fig. 9 does show N=5/11/100, but those are simulations of Eq. (5), not LLM agents. Consequently the distinctive, falsifiable prediction of the theory--N dependence of the rounding--is never tested with actual LLMs. The observed |m(R)| vs. T,b phase diagrams are also exactly what a single stochastic LLM with a small logit bias would produce: at low T, argmax/softmax selection over the biased logits makes the choice nearly deterministic, driving |m|~1 even with no interaction. Without an N sweep or a no-debate control, 'collective norm emerges' cannot be separated from 'a single biased LLM is deterministic at low T.' This matters because the abstract and Section 5 claim the threshold lambda rho N/T ~ 1 and finite-N rounding, not merely low-temperature amplification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies biased consensus in multi-agent LLM debates. The authors first show, in preliminary experiments on investment recommendation and LLM-as-a-Judge, that low sampling temperature leads to rapid convergence to a biased collective decision. They then propose a mean-field social-dynamics model (Eq. 5) in which each agent's choice is sampled from a Boltzmann distribution over an effective field combining conformity, neutral field, and bias; the model predicts a phase transition to collective bias when the conformity-to-noise ratio crosses a threshold, rounded by finite-N noise into a crossover. Controlled synthetic experiments with 11 LLMs on binary-choice and implicit-bias tasks produce phase diagrams in (T, b) that qualitatively match the predicted crossover; additional experiments show that sparsity, higher temperature, confidence visibility, and agent heterogeneity suppress the norm. The paper closes with real-world experiments showing that heterogeneous temperature mixing reduces bias and improves performance.","tokens_in":27803,"tokens_out":7641,"duration_ms":80780,"significance":"If the central claim holds, the paper makes a significant contribution: it offers a compact, interpretable account of a safety-relevant phenomenon in multi-agent LLM systems, with actionable design levers (temperature, sparsity, heterogeneity). The derivation in Appendix B is standard and clean, the code is released, and the breadth of LLMs (11 models, 5 families) plus two real-world tasks strengthens the empirical contribution. The finite-size rounding prediction is a genuinely falsifiable target, and the heterogeneity result is a useful theory-grounded intervention. However, as detailed below, the validation as it stands is partly circular and, critically, does not yet test the distinctive finite-size prediction with actual LLMs.","major_comments":[{"comment":"The paper's central claim is that the observed crossover is finite-size rounding of a phase transition, with noise O(1/sqrt(rho N)) in Eq. (5); yet no LLM experiment varies the number of agents N. All synthetic LLM debates use N=11 (Fig. 2 and Fig. 10), and the real-world tasks use N=6 and N=10. Figure 9 shows the N-dependence of the crossover only for numerical simulation of Eq. (5), not for LLM agents. Consequently the observed |m(R)| vs. T and b phase diagrams are equally consistent with a single stochastic LLM whose softmax over a logit-biased distribution becomes deterministic at low T, with no interaction among agents. Without an N-sweep (e.g., N=2, 5, 11, 30 with otherwise identical prompts) or a no-debate control in which agents do not see others' answers, the specific finite-size scaling content of the theory (sharpening with N, threshold shift toward lambda rho N/T ~ 1) is unvalidated. This is load-bearing because the abstract and Section 5 claim finite-size rounding, not merely low-temperature amplification.","section":"Section 6.1 and Fig. 9"},{"comment":"The quantitative 'theoretical predictions' are in-sample fits. The parameters lambda and h_bias (and the noise constant c and logit sensitivity w_b) are estimated by fitting the one-step map (Eq. (5)) to the same debate trajectories that are later compared with the predicted curves; Fig. 11's caption explicitly describes the curves as coming from the 'fitted mean-field model.' Agreement between theory and experiment under these conditions is a consistency check on the assumed functional form, not an out-of-sample validation of the predicted phase boundary lambda rho N/T ~ 1. The authors should either hold out part of the (T, b) grid or the round dimension when fitting and then predict the held-out data, or report the fitted parameter values and their uncertainties so the reader can assess whether the threshold location is genuinely predicted rather than absorbed by free parameters.","section":"Section 6.1.2, Eq. (21), and Figs. 3, 4, 11"},{"comment":"The model assumes each LLM agent's update probability has the exact Boltzmann form P(sigma_i(t+1)=sigma') proportional to exp(-H_i(sigma')/T) with constant conformity lambda and bias h_bias. This is a strong assumption about how autoregressive token-level generation maps to discrete-choice probabilities; the paper provides no independent check (e.g., a goodness-of-fit test of the implied logistic regression on conditional choice data, or an invariance test of the inferred lambda across T and b). Because lambda and h_bias are the quantities that determine the predicted threshold lambda rho N/T ~ 1, the quantitative transfer of the phase diagram to actual LLM systems rests on this unvalidated mapping. I regard this as a correctness-risk concern rather than a fatal flaw, but it needs to be addressed, or the claims weakened to qualitative phenomenology, before the quantitative predictions are presented as tested.","section":"Section 5, Eqs. (2)-(4), and Appendix B"},{"comment":"The real-world experiments use a different model than the core Eq. (5), adding performance feedback alpha R_i(t-1) and task-aligned correctness g_i(t-1) with parameters (J, alpha, b) estimated by regression on the same logs. The claim that heterogeneity 'reduces bias and improves performance' is supported by effect sizes that appear modest relative to the reported SEM (e.g., bias norm 0.8979 +/- 0.0096 vs. 0.9177 +/- 0.0096 in investment; performance 1.607 +/- 0.116 vs. 1.522 +/- 0.106), yet no significance test or confidence interval for the difference is reported. Please provide a statistical test (or at least a bootstrap CI) for the heterogeneity improvement, and clarify whether the real-world model is a separate fit or a specialization of Eq. (6).","section":"Section 6.2 and Appendix C.3.1, Eqs. (22)-(23)"}],"minor_comments":[{"comment":"The notation 'hi' in the caption should be made consistent with h_bias(sigma_i) used in Eq. (4); the caption also would benefit from explicit definitions of the plotted bias and conformity variables.","section":"Figure 5 caption"},{"comment":"The caption appears to contain unrendered placeholders ('x3, x3, and x3' and '( , , and )'); please ensure the model symbols are properly typeset in the final version.","section":"Figure 6 caption"},{"comment":"The caption refers to 'Eq. 13' for the stochastic mean-field dynamics; Eq. (13) gives the Fokker-Planck coefficients, not the recursion used for predictions (Eqs. (5)/(19)). The cross-reference should be corrected.","section":"Figure 11 caption"},{"comment":"The notation p_a(t) in Eq. (16) is not distinguished from m_a(t) in Eq. (17); please define explicitly that p_a is the single-agent state probability and m_a is the population fraction, to avoid confusion in the q>=3 extension.","section":"Appendix B, Eqs. (16) and (17)"},{"comment":"The sentence 'In both tasks, we systematically manipulate the token-level logit bias' could be misleading because the Implicit Bias task also has a semantic gender bias; clarify that the external, experimenter-controlled component is the logit bias on one option (e.g., Jane), as described in Appendix C.1.1.","section":"Section 6.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for ICML and likely to attract attention. The main risk is the reviewer concern about the missing N-sweep and the in-sample validation; I would suggest the editor ask for a revised version that includes at least one LLM N-sweep or an explicit no-debate control, and out-of-sample validation of the theoretical predictions. Also note that the paper cites Ashery et al. (2025) but does not compare to it in detail; the novelty relative to that work should be clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The quick take: this paper gives multi-agent LLM debates a clean mean-field treatment—Curie-Weiss/Brock-Durlauf with sampling temperature as noise, logit bias as a field, sparsity, and finite-N fluctuations—and shows that the predicted crossover appears in real LLM debates. That framing is new for this application and the derivation in Appendix B is standard and clear. The synthetic phase diagrams (Figs. 2a,b and 9) match the theory, and the heterogeneity-smoothing prediction in Fig. 6 is a nice, genuine use of Eq. (6).\n\nBut the stress-test note lands. The paper's headline claim is finite-size rounding of a phase transition, and that claim is never tested with LLMs. Fig. 9 varies N=5/11/100, but only for simulations of Eq. (5). The synthetic LLM debates use N=11, and the real tasks use N=6 and N=10. There is no debate/no-interaction control. Without an N sweep or a single-agent control, the observed temperature crossover is exactly what a single biased LLM does at low T: softmax/argmax selection over biased logits drives |m| close to 1 even with zero conformity. So \"collective norm emerges\" and \"finite-N rounding\" are not yet supported by the LLM data. This is the main weakness, and it is fixable—run the same tasks at, say, N=3, 5, 11, 21 and add a no-debate baseline.\n\nSecond, several \"theoretical predictions\" are in-sample. The curves in Figs. 3, 4, and 11 fit λ and h_bias to the same trajectories (Eq. 21) and then compare with those trajectories. The threshold criterion λρN/T ≳ 1 comes from the model, so the qualitative shape is not circular, but the word \"prediction\" oversells the quantitative agreement. A split-sample or leave-one-out fit would help.\n\nThird, the Gibbs-form mapping—choice probability proportional to exp(-H/T) with static H—is asserted rather than validated. Real LLM outputs are autoregressive token samples. The fitting procedure can absorb this mismatch, so the model works as an effective theory, but the quantitative threshold should not be treated as directly transferable to arbitrary LLM systems until that mapping is checked.\n\nMinor point: the real-world heterogeneity gains are modest and often within a couple of standard errors; the paper's phrasing \"reduce bias and improve performance\" is a bit strong.\n\nWho this is for: AI safety researchers and anyone using spin models for collective LLM behavior. It deserves peer review—the gaps are addressable with a proper N sweep, a single-agent control, and out-of-sample fits. I'd send it out, with those requests attached.","headline":"Clean mean-field theory for LLM debate bias, but the finite-N rounding claim rests on simulations only; the LLM experiments never vary N.","tokens_in":28314,"tokens_out":2848,"would_cite":true,"duration_ms":28808,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In multi-agent LLM debates, low sampling noise turns small individual biases into a near-unanimous biased consensus.","keywords":["multi-agent LLM debate","collective bias","biased consensus","phase transition","finite-size crossover","mean-field spin model","sampling temperature","LLM-as-a-judge"],"falsifier":"Run a debate with a fixed majority: fix temperature and logit bias, then measure the empirical probability that an agent chooses +1 as a function of the previous round's collective norm $m(t)$. If that curve is not the predicted logistic (tanh) form—for example, if it is non-monotonic or its slope depends on prompt wording rather than only on $\\lambda\\rho N/T$—then the Gibbs mapping fails and the phase-transition explanation does not transfer. Alternatively, measure the crossover width in $|m(R)|$ across $N=5,10,100$ and check whether it shrinks like $1/\\sqrt{\\rho N}$; if it does not, the finite-size rounding story is wrong.","tokens_in":1906,"feed_emoji":"⚖️","tokens_out":8858,"duration_ms":127891,"temperature":0.7,"pith_summary":"Multi-agent LLM debates can fail because of collective bias: when the sampling temperature is low, even a small bias in individual LLMs gets amplified into a near-unanimous norm, often a biased one. The paper models each agent's discrete choice as a softmax (Boltzmann) decision over an energy that combines conformity to other agents with an individual bias, and derives a mean-field recursion for the group alignment $m(t)$. The model predicts a phase transition: biased consensus emerges whenever the combination $\\lambda\\rho N/T$ (conformity times interaction density times group size divided by temperature) crosses about one. Because real systems have small $N$, the sharp transition is rounded into a smooth crossover, and experiments across several LLM families reproduce the predicted dependence on temperature and bias. If correct, this means the debate protocol itself can turn small biases into hard-to-reverse group decisions.","feed_headline":"LLM debates lock into biased consensus below a noise threshold","feed_subtitle":"A spin-model threshold shows how a few LLM agents amplify weak individual bias into a locked group norm.","key_machinery":"The carrying object is a mean-field Ising-type recursion for the collective alignment, $m(t+1) = \\tanh\\left(\\frac{1}{T}(\\lambda\\bar z m(t)+h\\pm\\gamma/2)\\right)+\\eta(t)$, with finite-size noise $\\eta(t)=O(1/\\sqrt{\\rho N})$. Each LLM agent's discrete choice is sampled from the softmax distribution over an effective energy that penalizes disagreement with other agents (conformity $\\lambda$), adds a task-aligned prior $h$, and adds a small bias $\\gamma$ favoring a subset of options; $\\bar z\\simeq\\rho N$ is the effective number of neighbors under interaction sparsity $\\rho$. This recursion converts the microscopic update rule into a macroscopic map whose fixed-point structure predicts the phase-transition threshold $\\lambda\\bar z/T\\gtrsim 1$, and finite-size noise is what rounds the transition into the observed crossover.","core_discovery":"The central claim, stated in Section 5, is that a collective norm emerges once conformity and noise cross a critical threshold, $\\lambda\\bar z/T \\gtrsim 1$ (equivalently $\\lambda\\rho N/T \\gtrsim 1$), so that low temperature and weak individual bias produce a near-saturated biased consensus $|m(R)|\\approx 1$ in multi-agent LLM debates. The author derives this from a mean-field spin model with a Boltzmann update rule, treats sampling temperature as inverse noise $\\beta=1/T$, and shows that the thermodynamic-limit phase transition is rounded into a finite-size crossover for the small agent numbers used in practice. Controlled experiments on binary and implicit-bias tasks across several LLM families find the same crossover structure and reproduce the predicted dynamics, and heterogeneous agent mixtures are shown to smooth the response, weakening biased lock-in and improving performance on investment and LLM-as-a-judge tasks.","pith_inferences":["A direct test of the load-bearing Gibbs assumption would fit the logistic curve for choices versus majority opinion; if the slope depends on prompt wording or non-monotonicity appears, the quantitative threshold $\\lambda\\rho N/T\\gtrsim 1$ would not transfer to real LLM systems.","The same rounded-transition mechanism is likely to appear in any multi-agent protocol whose agents sample from softmax preferences over other agents' outputs—committee reviews, market simulations, or ensemble generalists—so the temperature knob may be a general safety lever.","An easy extension would measure the crossover width in $|m(R)|$ across group sizes $N=5,10,100,1000$; the theory predicts it shrink like $1/\\sqrt{\\rho N}$, which would confirm the phase-transition interpretation rather than a mere prompting artifact.","The paper's diagnostic fits suggest a practical deployment check: estimate each LLM's conformity $\\lambda$ and bias $\\gamma$ from a small debate trace, then place the ensemble on the phase diagram before launch."],"forward_implications":["Raising sampling temperature suppresses biased consensus in both investment recommendations and LLM-as-a-judge, at least within the tested range.","Reducing effective interaction—sparser communication graphs, lower sycophancy prompting, or fewer observed neighbors—moves the system below threshold and weakens biased lock-in.","Mixing agents with heterogeneous sampling temperatures smooths the collective response and reduces the bias norm while preserving or improving task performance.","For multi-choice ($q\\ge 3$) debates, the same mechanism holds with a Potts-type update replacing the binary tanh recursion.","Fitting debate trajectories to the mean-field map recovers per-model conformity and bias, making the framework usable as a diagnostic before deployment."],"supporting_citations":[{"why":"Supplies the multi-agent debate protocol and the independent-initial-decision procedure used throughout the experiments.","marker":"Du et al., 2023"},{"why":"Provides the discrete-choice-with-social-interactions model that grounds the spin-model formulation for interacting populations.","marker":"Brock & Durlauf, 2001"},{"why":"Gives the finite-size scaling result that justifies rounding the sharp transition into a crossover for small N.","marker":"Privman & Fisher, 1984"},{"why":"Supplies the implicit gender bias dataset used for the controlled synthetic task.","marker":"Borah & Mihalcea, 2024"},{"why":"Provides the investment recommendation setup and the US/technology concentration bias metric used in the realistic task.","marker":"Winder et al., 2025"},{"why":"Supplies the MT-Bench data and human ground-truth labels used in the LLM-as-a-judge task.","marker":"Zeng et al., 2024"},{"why":"Establishes the sycophancy and LLM-as-judge bias concepts that motivate the conformity parameter lambda.","marker":"Zheng et al., 2023"},{"why":"Empirically documents group conformity in multi-agent systems, giving the conformity term in the effective energy its behavioral grounding.","marker":"Choi et al., 2025b"}],"fun_headline_variants":["Low noise locks LLM debates into biased consensus","Noise threshold triggers biased consensus in LLM debates","Heterogeneous agents suppress biased consensus in LLM debates","Collective bias emerges below critical noise in LLM debates","Phase transition to biased consensus in multi-agent LLM debates"],"cache_read_input_tokens":30464,"weakest_assumption_plain":"The argument assumes that an LLM's debate choice really is drawn from a softmax distribution over a fixed, static scalar energy, so that temperature works as inverse noise and conformity and bias stay constant; if autoregressive token sampling breaks that Gibbs form, the predicted threshold may not apply to real LLM systems.","fun_headline_variants_meta":{"raw":{"variants":["Low noise locks LLM debates into biased consensus","Noise threshold triggers biased consensus in LLM debates","Heterogeneous agents suppress biased consensus in LLM debates","Collective bias emerges below critical noise in LLM debates","Phase transition to biased consensus in multi-agent LLM debates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001134,"raw_usage":{"total_tokens":4684,"prompt_tokens":895,"completion_tokens":3789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":3711}},"tokens_in":511,"tokens_out":3789,"duration_ms":24508,"temperature":1.0,"reasoning_tokens":3711,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:58:32.175356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a debate with a fixed majority: fix temperature and logit bias, then measure the empirical probability that an agent chooses +1 as a function of the previous round's collective norm $m(t)$. If that curve is not the predicted logistic (tanh) form—for example, if it is non-monotonic or its slope depends on prompt wording rather than only on $\\lambda\\rho N/T$—then the Gibbs mapping fails and the phase-transition explanation does not transfer. Alternatively, measure the crossover width in $|m(R)|$ across $N=5,10,100$ and check whether it shrinks like $1/\\sqrt{\\rho N}$; if it does not, the finite-size rounding story is wrong.","supporting_citations":[],"review_version":2}