Pith. sign in

REVIEW 4 major objections 5 minor 21 references

Emergence of Biased Consensus in Multi-Agent LLM Debates

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read In multi-agent LLM debates, low sampling noise turns small individual biases into a near-unanimous biased consensus.

desk verdict Clean mean-field theory for LLM debate bias, but the finite-N rounding claim rests on simulations only; the LLM experiments never vary N. read the letter →

arxiv 2608.02827 v1 pith:6ICAORHG submitted 2026-08-03 cs.MA

Maya Okawa This is my paper · ORCID
classification cs.MA
keywords multi-agentLLMdebatecollectivebiasbiasedconsensusphasetransitionfinite-sizecrossovermean-fieldspinmodelsamplingtemperatureLLM-as-a-judge
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Section 6.1 and Fig. 9] 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.
  2. [Section 6.1.2, Eq. (21), and Figs. 3, 4, 11] 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.
  3. [Section 5, Eqs. (2)-(4), and Appendix B] 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.
  4. [Section 6.2 and Appendix C.3.1, Eqs. (22)-(23)] 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).
minor comments (5)
  1. [Figure 5 caption] 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.
  2. [Figure 6 caption] 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.
  3. [Figure 11 caption] 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.
  4. [Appendix B, Eqs. (16) and (17)] 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.
  5. [Section 6.1.1] 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.

Circularity Check

1 steps flagged · score 6.0 of 10

Theoretical curves used to validate Eq. (5) are fitted to the same one-step transitions they are compared with; the central phase-transition phenomenology is otherwise a self-contained mean-field derivation.

  1. fitted input called prediction [Section 6.1.2 ('Theory as a Diagnostic Tool', Fig. 3) and Appendix C.1.1 (Fitting Details, Eq. (21))]
    "For the theoretical predictions, we fit the one-step transition m(t)→m(t+1) in the debate trace to Eq. (5) (see Appendix C.1.1 for fitting details). ... we fit the mean-field one-step map in Eq. (5) by nonlinear least squares. ... We estimate parameters by minimizing the prediction error over all observed transitions, min_{λ,h^bias} Σ_t (m(t+1)−m̂(t+1))^2."

    The 'theoretical predictions' drawn in Figs. 3, 4(a,b), 6(a), and 11 are generated by iterating Eq. (5)/(19) with λ and h_bias estimated by least squares on the very debate trajectories whose final consensus |m(R)| is then plotted as 'Empirical vs Theory.' The fitted data include the final-round transitions, so the agreement displayed is an in-sample measure of the fit, not an independent out-of-sample confirmation of the predicted phase transition or threshold λρN/T≳1. The one-step map does add dynamical content beyond directly fitting the final consensus, so the circularity is partial; the heterogeneity prediction of Eq. (6) is genuinely out-of-sample because parameters are fit only on homogeneous populations.

full rationale

The analytic part is not circular: Eq. (5) is derived in Appendix B from the assumed Boltzmann/softmax update Eq. (2) with a mean-field interaction, and no load-bearing self-citation or imported uniqueness theorem is used. The phase-transition threshold λρN/T≳1 follows from the deterministic map, and the finite-N noise estimate η=O(1/√ρN) is derived from the random-sequential update. The circularity is located in the empirical validation stage: the paper labels as 'theoretical predictions' curves that are generated from parameters fitted to the same one-step transitions used for comparison (Eq. (21)), so the agreement in Figs. 3, 4, 6(a), and 11 is partly a fit diagnostic. The heterogeneity experiment is a valid out-of-sample prediction. The absence of an LLM N-sweep weakens the finite-N rounding claim, but that is a missing-evidence/correctness issue, not circularity. Overall partial circularity in the empirical support warrants 6 rather than a lower score.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the mean-field reduction, the Boltzmann choice model, a static bias field, and standard finite-size scaling results; the quantitative comparison to data additionally depends on parameters estimated from the same debate traces.

free parameters (5)
  • conformity parameter λ_i = per-LLM values estimated from debate trajectories (Fig. 5, Fig. 12)
    Fitted by nonlinear least squares to one-step transitions (Eq. 21); used to generate theoretical curves in Figs. 3, 4, and 11.
  • local bias field h_bias = per-LLM values estimated from debate trajectories (Fig. 5)
    Fitted together with λ; represents the systematic preference in the effective field Eq. (4).
  • logit-bias sensitivity w_b = not reported
    In Eq. (19), the controlled logit bias b enters the field as w_b b; w_b must be inferred to produce predictions from the binary-choice data.
  • finite-size noise constant c = not reported
    In Appendix C.1.1, the noise variance is set to σ^2 = c^2/N; c is chosen by hand to absorb discretization effects, adding a free constant to the predicted curves.
  • real-world model parameters (J, α, b) = estimated via regression (Appendix C.3.1)
    The investment and LLM-as-judge analyses rely on regression-fitted social influence, performance feedback, and bias parameters to support the heterogeneity claims.
assumptions (4)
  • domain assumption Mean-field approximation: the local social field is approximated by ρN*m(t) (Appendix B, Eq. 5 derivation).
    Used to reduce the microscopic spin dynamics to the scalar recursion Eq. (5); standard but uncontrolled for sparse, heterogeneous, or small networks.
  • domain assumption LLM debate choices follow the Boltzmann (softmax) update in Eq. (2) with β=1/T (Section 5, Update Rule).
    This maps LLM temperature to inverse noise and discrete debate answers to spin states; not independently validated.
  • domain assumption Individual bias is a static additive field h_bias(σ)=γI(σ in S) (Section 5, Eq. 4).
    The bias is assumed constant across rounds and independent of the content of other agents' responses.
  • standard math Standard statistical-physics results: mean-field Curie-Weiss transition at βJ=1 and finite-size rounding into a crossover (Privman & Fisher 1984).
    Used to interpret the observed temperature dependence as a rounded phase transition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Emergence of Biased Consensus in Multi-Agent LLM Debates." pith.science (2026). https://pith.science/paper/6ICAORHG

@misc{pith2026260802827,
  author       = {Pith},
  title        = {Pith review of: Emergence of Biased Consensus in Multi-Agent LLM Debates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ICAORHG}},
  note         = {Machine review of arXiv:2608.02827}
}
read the original abstract

Multi-agent LLM debates achieve strong performance on decision-making tasks as well as problem-solving benchmarks, yet their safety and fairness risks remain poorly understood. Notably, interaction can amplify the biases of single LLMs, raising concerns for real-world deployment. We identify the emergence of collective (often biased) norms in multi-agent LLM debates and show that noise (e.g., LLM sampling temperature) is a key driver. To explain this, we propose an analytical framework drawing on physics-inspired theoretical models of social dynamics. We predict a phase transition to collective bias when conformity surpasses a critical threshold given the LLMs' initial bias and debate noise. We test the theoretical predictions through controlled experiments and observe a finite-size crossover consistent with an underlying phase transition. We further find that agent heterogeneity suppresses emergence by smoothing (rounding) this transition. Finally, we show that these insights generalize to realistic decision-making tasks, including investment decisions and LLM-as-a-judge evaluation.

Figures

Figures reproduced from arXiv: 2608.02827 by the authors.

Figure 1
Figure 1. Biased consensus can emerge from smaller initial biases, and sampling temperature is a key knob. Bias and per￾formance in multi-agent LLM debates (Du et al., 2023) on realistic tasks: (a) investment recommendations (bias: U.S. concentration; performance: return) and (b) LLM-as-a-judge on MT-Bench data (bias: self-bias; performance: agreement with human ground truth labels). Left: final-round bias (pink) and performa… view at source ↗
Figure 2
Figure 2. Under conformity, collective biased norms emerge when single-LLM bias and stochasticity reach a critical balance, yielding |m(R)| ∼ 1. (a,b) Mean-field predictions from Eq. (5) for (a) N = 1000 and (b) N = 11 LLM agents. (c–e) Empirical phase diagrams of the collective norm |m(R)| as a function of single-LLM bias b and sampling temperature T: (c) GPT-4 Nano on Binary Choice task (O vs. I); (d) GPT-4 Nano on Implicit… view at source ↗
Figure 3
Figure 3. Debates quickly lock into biased consensus at low T, consistent with theory. Empirical (solid) and theoretical (dashed) collective norm for Llama-405B on the Implicit Bias task versus (a) sampling temperature T and (b) round t. Shaded areas denote SEM. Our theoretical model reproduces the dynamics: at low T, moderate local bias leads to rapid convergence to near-complete consensus |m(R)| ∼ 1 within a few rounds. 6. … view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Our theoretical framework quantifies conformity and bias. Estimated gender bias hi and conformity λi for each LLM. Each point corresponds to one LLM, and marker size indicates the number of LLM parameters. Parameters are inferred by fitting the one-step dynamics m(t)→m…
Figure 6
Figure 6. Figure 6: Mixing LLM types suppresses biased norms, as predicted by the theory. Final-round consensus m(R) versus bias field b for three LLM: (a) Homogeneous populations ( ×3, ×3, and ×3); (b) Heterogeneous populations ( , , and ). Shaded areas indicate SEM over 20 runs. our for…
Figure 8
Figure 8. Figure 8: (a) Technology-sector bias in investment recommendation and (b) Self bias in GPT-4 for the LLM-as-a-judge task on MT-Bench data. Left panels show bias and task performance as a function of sampling temperature T. Right panels show how bias evolves over debate rounds. E…
Figure 9
Figure 9. Figure 9: Final collective norm |m(R)| as a function of the local bias h(σi = +1) and the effective interaction strength βλiJij (t). (a) Theoretical prediction computed from the stationary distribution of a one-dimensional Fokker–Planck equation approximating the finite-size mea…
Figure 10
Figure 10. Figure 10: Collective norm |m(t)| as a function of single-LLM bias b and sampling temperature T. The leftmost panel shows the mean-field theoretical prediction. Remaining panels report empirical results for GPT-4.1 Nano, Llama-405B, and DeepSeek-V3 on two tasks: Binary Choice ta…
Figure 11
Figure 11. Figure 11: Empirical consensus (solid lines with error bars) and mean-field theoretical predictions (dashed lines; dotted lines indicate the noise term) at the final round t = R for homogeneous groups of N = 11 LLM agents, shown as a function of the local bias b at fixed tempera…
Figure 12
Figure 12. Figure 12: Inferred token bias and interaction strength on the token-choice task. We report the intrinsic token bias h(σi = +1) and interaction strength λiJij (t) inferred from debate trajectories on a binary token-choice task. Compared to the gender bias task in [PITH_FULL_IMA…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 6 canonical work pages

  1. [1]

    Begin with a single, concise explanation (1-2 sentences)

    work page

  2. [2]

    Investment Type

    Then output ONLY a Python list of dictionaries in this exact structure: [ {"Investment Type": ..., "Name": ..., "Ticker Symbol": ..., "Amount to Invest": ...}, ... ] During intermediate rounds fromt= 1tot=r, the instruction is updated as follows: These are the portfolio proposals from others: <a list of other agents’ answers> Revise and improve your portf...

    work page 2025

  3. [9]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl-long

    doi:10.18653/v1/2025.acl-long 2025

  4. [12]

    ISBN 979-8-89176-332-6

    Association for Computational Linguistics. ISBN 979-8-89176-332-6. doi: 10.18653/v1/2025.emnlp-main

    doi:10.18653/v1/2025.emnlp-main 2025

  5. [14]

    URL https: //aclanthology.org/2025.gebnlp-1.20/

    doi: 10.18653/v1/2025.gebnlp-1.20. URL https: //aclanthology.org/2025.gebnlp-1.20/. Stengel-Eskin, E., Hase, P., and Bansal, M. Teaching mod- els to balance resisting and accepting persuasion. In Chiruzzo, L., Ritter, A., and Wang, L. (eds.),Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Ling...

    doi:10.18653/v1/2025.gebnlp-1.20 2025

  6. [15]

    ISBN 979-8-89176-189-6

    Association for Computational Linguistics. ISBN 979-8-89176-189-6. doi: 10.18653/v1/2025.naacl-long

    doi:10.18653/v1/2025.naacl-long 2025

  7. [22]

    +” favors +1 and “−

    =ρ∈(0,1] (thusρ= 1 recovers the fully-connected case). Letki =P jJij be the degree of agenti and ¯z=E[ki] = ρ(N−1)≃ρN the mean degree. Under the mean-field approximation, the local social field satisfies P jJijσj(t)≈ kim(t)≈¯zm(t) . We instantiate the interaction potential as ϕ(σ,σ′) =σσ ′ and restrict the (neutral) external field to the linear binary for...

    work page 2025

  8. [101]

    findings-naacl.101/

    URL https://aclanthology.org/2025. findings-naacl.101/. Du, Y ., Li, S., Torralba, A., Tenenbaum, J. B., and Mor- datch, I. Improving factuality and reasoning in lan- guage models through multiagent debate.arXiv preprint arXiv:2305.14325, 2023. Dudy, S., Tholeti, T., Ramachandranpillai, R., Ali, M., Li, T. J.-J., and Baeza-Yates, R. Unequal opportunities:...

    arXiv 2025

Show all 21 references
  1. [184]

    findings-emnlp.184/

    URL https://aclanthology.org/2024. findings-emnlp.184/. Weidlich, W. Physics and social science — the approach of synergetics.Physics Reports, 204(1):1–163, 1991. ISSN 0370-1573. doi: https://doi.org/10.1016/0370-1573(91)90024-G. URL https://www.sciencedirect.com/ science/arti...

    2024

  2. [211]

    findings-naacl.211/

    URL https://aclanthology.org/2024. findings-naacl.211/. Cisneros-Velarde, P. Biases in opinion dynamics in multi- agent systems of large language models: A case study on funding allocation. In Chiruzzo, L., Ritter, A., and Wang, L. (eds.),Findings of the Association for Comput...

    2024 doi

  3. [264]

    doi: 10.1201/ 9781003278290-37

    Auerbach Publications, 2022. doi: 10.1201/ 9781003278290-37. URL https://doi.org/10. 1201/9781003278290-37. Armstrong, L., Liu, A., MacNeil, S., and Metaxa, D. The Silicon Ceiling: Auditing GPT’s race and gender biases in hiring. InProceedings of the 4th ACM Conference on Equi...

    2022

  4. [328]

    acl-long.328/

    URL https://aclanthology.org/2025. acl-long.328/. Flache, A., M¨as, M., Feliciani, T., Chattoe-Brown, E., Def- fuant, G., Huet, S., and Lorenz, J. Models of social influence: Towards the next frontiers.Journal of Ar- tificial Societies and Social Simulation, 20(4):2, 2017. ISS...

    2025

  5. [412]

    naacl-long.412/

    URL https://aclanthology.org/2025. naacl-long.412/. Takata, R., Masumori, A., and Ikegami, T. Emergent social dynamics of LLM agents in the El Farol bar problem. arXiv preprint arXiv:2509.04537, 2025. Tao, Y ., Viberg, O., Baker, R. S., and Kizilcec, R. F. Cultural bias and cu...

    2025

  6. [427]

    findings-emnlp.427/

    URL https://aclanthology.org/2024. findings-emnlp.427/. Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y ., Wang, R., Yang, Y ., Tu, Z., and Shi, S. Encouraging divergent thinking in Large Language Models through multi-agent debate.arXiv preprint arXiv:2305.19118, 2023. Liu, Y ....

    2024 arXiv

  7. [511]

    acl-long.511/

    URL https://aclanthology.org/2024. acl-long.511/. Wang, Z., Wu, Z., Guan, X., Thaler, M., Koshiyama, A., Lu, S., Beepath, S., Ertekin, E., and Perez-Ortiz, M. JobFair: A framework for benchmarking gen- der hiring bias in large language models. In Al- Onaizan, Y ., Bansal, M., ...

    2024 doi

  8. [545]

    findings-emnlp.545/

    URL https://aclanthology.org/2024. findings-emnlp.545/. Brock, W. A. and Durlauf, S. N. Discrete choice with social interactions.The Review of Economic Studies, 68(2):235–260, 2001. doi: 10.1111/1467-937X.00168. URL https://doi.org/10.1111/1467-937X. 00168. Buyl, M., Rogiers, ...

    2024 doi arXiv

  9. [606]

    findings-acl.606/

    URL https://aclanthology.org/2025. findings-acl.606/. Ki, D., Rudinger, R., Zhou, T., and Carpuat, M. Multiple LLM agents debate for equitable cultural alignment. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.),Proceedings of the 63rd Annual Meeting of the Ass...

    2025

  10. [650]

    acl-long.650/

    URL https://aclanthology.org/2026. acl-long.650/. Choi, M., Kim, K., Chae, S., and Baek, S. An empirical study of group conformity in multi-agent systems. In Findings of the Association for Computational Linguis- tics: ACL 2025, pp. 5123–5139, 2025b. Chuang, Y .-S., Goyal, A.,...

    2026

  11. [1210]

    acl-long.1210/

    URL https://aclanthology.org/2025. acl-long.1210/. Kim, Y ., Park, C., Jeong, H., Chan, Y . S., Xu, X., Mc- Duff, D., Lee, H., Ghassemi, M., Breazeal, C., and Park, H. W. MDAgents: An adaptive collaboration of LLMs for medical decision-making, 2024. URL https://arxiv.org/abs/2...

    2025 arXiv

  12. [1653]

    emnlp-main.1653/

    URL https://aclanthology.org/2025. emnlp-main.1653/. Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. Generative agents: Interactive 11 Emergence of Biased Consensus in Multi-Agent LLM Debates simulacra of human behavior. InIn the 36th An...

    2025 arXiv

  13. [2023]

    Zhou, X., Zhu, H., Mathur, L., Zhang, R., Qi, Z., Yu, H., Morency, L.-P., Bisk, Y ., Fried, D., Neubig, G., and Sap, M

    Curran Associates Inc. Zhou, X., Zhu, H., Mathur, L., Zhang, R., Qi, Z., Yu, H., Morency, L.-P., Bisk, Y ., Fried, D., Neubig, G., and Sap, M. SOTOPIA: Interactive evaluation for social intelli- gence in language agents. 2024. 13 Emergence of Biased Consensus in Multi-Agent LL...

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.