acceptodds
Under review as a conference paper at ICLR 2027

How does Adversarial Influence Scale in Multi-Agent Systems?

Abstract

Multi-agent deliberation can improve performance, but what happens when some agents do not act in good faith? In practice, an agent may be deceptive and work to subvert the group, whether through its own objectives or external instruction. We study how susceptibility to deception scales as groups increase in size and deceivers become more prevalent. It is not the number of agents in the group that matters, but the proportion of deceivers. We observe that the defection rate, how often initially correct agents switch to an incorrect final answer, rises linearly with this proportion. Whereas humans in comparable conformity studies are reliably swayed only when misleading confederates form a majority, LLM agents defect regularly even when deceivers remain a minority. Susceptibility also depends on which models are interacting, especially on the honest agent side. Unexpectedly, allowing deceivers to coordinate privately can make them less effective. Altogether, our results show that adding more agents is therefore not a sufficient defense, because the adversary can simply scale with the group.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, Safety, Multi-Agent, Evaluation, Cognitive Science

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rwZqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed to experiment on how the proportion of deceive agents affects the final voting results of a multi-agent system and it found that defection rate rises linearly with the proportion. The task is to ask the agents to first answer questions from HLE, then discuss publicly and finally do the votes. The authors experimented on the proportion of 0, 1/5, 1/3 and 3/7 and the number of agents varied in 2, 4, 8, 12. Also the authors statistically analyzed some reasons behind the results, such as the rate of citing of the wrong interpretation, the rate of redirecting faithful agents' arguments, etc.

Critical strengths
  1. The paper was inspired by the human conformity studies and experimented on LLM agents, and finally got some unexpected results..
Critical weaknesses
  1. Contribution. I think the contribution of this paper is slightly limited. It only discusses the effect under deception, what about other scenarios? For example, what would the proportion of malicious agents have effect on the final results in the scenario of safety? And the social effect of other agents is actually discussed earlier (eg. BlockAgents: Towards Byzantine-Robust LLM-Based Multi-Agent Coordination via Blockchain), so just giving the results of "the proportion of deceive agents linearly effects the defection rate" seems somewhat insufficient.

  2. Comprehensiveness. The number of experimented agents is at most 12, and the experiments are only done on 4 scales (2, 4, 8, 12), which is not sufficiently enough. And the experimented proportions are only 4 (0, 1/5, 1/3, 3/7). I believe if experimenting on the proportion larger than 3/7, then the linear relation may be not applicable, it may saturate to 1. What's more, I hope the experiments could be done on frontier LLMs, like gpt series and claude series. Besides, the ablation experiments of the system prompts are not done. I believe the system prompts giving to the deceptive agents and faithful agents will have effects on the results, but the authors fixed them instead. For example, how the prompts expressed with some kind of mood have effects on the results? Finally, the experiment details lack the params of calling the backend LLMs, such as the temperature, reasoning effort, etc.

Questions for authors

See weakness.

8 comments
AuthorsOpenReview ↗

Response to reviewer

We thank the reviewer for their thoughtful comments! We address each concern below.

W1: Contribution and broader relevance

Broader safety implications. Our study examines a fundamental mechanism underlying multi-agent safety risks: malicious agents persuading honest agents to reach incorrect conclusions. This form of deception can enable broader failures, including sabotage, misinformation, and unsafe decisions. We deliberately use HLE because its verifiable answers give honest agents a strong basis for defending correct conclusions. Yet even in this setting, deceptive minorities regularly persuade honest agents to abandon correct answers. In more ambiguous settings, such as safety assessments, determining what is correct may be harder, potentially increasing vulnerability to deception. Our findings therefore establish a foundational vulnerability relevant to broader multi-agent safety, while extending the experiments to these settings remains important future work.

Distinction from BlockAgents. BlockAgents proposes blockchain-based consensus against Byzantine agents; our work instead studies how persuasive adversarial influence scales with the number and proportion of deceivers. Crucially, deceivers can change honest agents' votes themselves. At a 3/7 deceiver proportion, the honest agents' majority reached the correct answer 13–17 percentage points less often (Gemini, Grok, Muse Glimmer), even after excluding all deceiver votes. Thus, deception can undermine the honest majority on which consensus-based defenses rely. We will cite BlockAgents in related work.

W2: Experimental comprehensiveness

Group sizes. Our groups contain up to 21 agents, not 12. The values 2, 4, 8, and 12 refer to honest agents. We evaluate 12 group compositions spanning 2–21 total agents (Table 1), across approximately 8,300 deliberations. This design varies group size at fixed deceiver proportions.

Proportions above 3/7. We agree that linearity need not extend beyond our tested range. We deliberately study groups with an honest majority: at a deceiver proportion of 1/2 or higher, deceivers can prevent a correct majority without persuading anyone. Our question is whether an honest majority protects against deception; our results show it does not. We find no evidence of saturation within the tested range: at 3/7, defection is still 35–52%, and a square-root curve allowing tapering fits worse than a straight line for all four models (Table 2). We make no claim about larger proportions.

GPT and Claude models. We evaluate four model families: Gemini 3.8 Flash, Grok 4.3, DeepSeek V4.1 Flash, and Muse Glimmer. Our study cost over $14,000.

Based on measured token usage under the same protocol, one additional 12-composition grid would cost approximately $20,000 with GPT-5.6 or $38,000 with Claude Opus 5. Either exceeds our entire study's cost. The consistent pattern across four families suggests it is not model-specific.

Prompt sensitivity and mood. We deliberately use open-ended prompts that specify each agent's objective without prescribing a particular mood, tone, or strategy. Our goal is to observe how agents naturally engage in deception, rather than impose specific behaviors. Indeed, deceivers spontaneously adapt their strategies during discussion, shifting from fabricated evidence toward conceding and redirecting, selective skepticism, and question reinterpretation (Figure 7). Prescribing a particular mood would constrain this behavior. We therefore hold prompts fixed to isolate the effect of group composition, while acknowledging that prompt variations may affect absolute defection rates.

Model parameters. Agents use provider-default temperature and reasoning settings. Answers are graded using GPT-5.4 and HLE's official grading instructions (Appendix C). We will add these details to Section 2.4.

In summary, across four model families and groups of 2–21 agents, honest defection increases approximately linearly with the proportion of deceivers, even when honest agents remain the majority. Increasing group size alone does not reliably protect against adversarial influence.

Sincerely,

The Authors

Thank you for the detailed response. I appreciate the clarifications regarding the experimental setup, particularly the group sizes and the focus on honest-majority settings. However, my concerns about the paper's contribution and experimental generalizability remain. While the observed relationship between deceiver proportion and honest-agent defection is interesting, the experiments are limited to a fixed prompting strategy and a single task setting (HLE). It remains unclear whether the observed approximately linear relationship holds under different prompting strategies or task environments. I understand the practical constraints on evaluating additional models, but I believe further evidence is needed to establish the broader applicability of the findings. Therefore, I will maintain my original rating.

AuthorsOpenReview ↗

Experiments to address generality (prompting, dataset)

We thank the reviewer for engaging with our response! To directly address the remaining concerns about generalizability, we conducted two new sets of experiments with Gemini 3.8 Flash: varying the instructions given to honest and deceptive agents and extending our evaluation to four additional benchmarks, including two safety-related benchmarks.

Prompting strategy and mood

We tested two alternative prompting styles, applied separately to deceivers and honest agents. The calm version preserved the original objective while removing emphatic language (capitalization, "must," and "never"). The emotional version introduced personal stakes ("This is extremely important to me. My career depends on the group choosing an [in]correct answer."). All prompts can be found in Appendix E in the attached revision.

We evaluated six group compositions: 8 honest agents with 0, 2, 4, or 6 deceivers (varying share), plus 4+3 and 12+9 at a fixed 3/7 deceiver share.

Across all five prompting conditions, defection increased significantly with deceiver proportion (p<0.001), following an approximately linear trend (R2>0.97), with no significant departure from linearity (p=0.85). The proportion model also significantly outperformed the count model (ΔAIC = 30.3; , p=0.02). Thus, all core results from the original paper still stand across prompting variations.

Task environment

We also evaluated four benchmarks that differ from HLE in task type, answer format, and safety relevance:

  • SimpleQA Verified: Short-answer factual recall without tools, testing whether deception remains effective without extended reasoning.
  • SuperGPQA-hard: Ten-option multiple-choice questions spanning 285 graduate disciplines, requiring deceivers to argue for existing incorrect options rather than invent answers.
  • WMDP: Biosecurity and cybersecurity questions, extending the evaluation to hazardous-knowledge domains.
  • SALAD-Bench: Safety judgments across six harm domains, requiring agents to distinguish safe from unsafe responses to harmful requests.

Together with HLE, these benchmarks cover factual recall, complex reasoning, multiple-choice decisions, and safety judgments.

Across all four additional benchmarks, defection increased significantly with deceiver proportion (permutation tests, all p<0.05), following approximately linear trends over the tested range (R2=0.86–0.95) with no significant departure from linearity (p = 0.15). The proportion model fit better than the count model on every benchmark and across all four benchmarks combined (ΔAIC = 21.3), while adding absolute deceiver count as a variable provided no significant improvement (p=0.32).

In summary, our central finding persists across five prompting conditions and five benchmarks across a variety of tasks: deceiver proportion consistently predicts adversarial influence better than absolute count, while increasing group size does not reliably protect the honest majority. These experiments directly address the reviewer's concerns about prompt sensitivity and generalizability beyond HLE. We will incorporate both sets of experiments into the appendix.

We sincerely thank the reviewer for suggesting that we test other prompting strategies and task environments. These experiments have made our findings substantially more robust and general, and the safety benchmarks make their real-world consequences far more salient: a deceptive minority can sway a group's judgments about hazardous knowledge and harmful requests.

We hope these additional results address the reviewer's remaining concerns, and we would be grateful if the reviewer would reconsider their assessment.

Sincerely,

The Authors

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

This paper studies how adversarial influence scales in MAS as the group size and proportion of deceptive agents increase. The authors evaluate four LLMs on HLE under different settings. The authors claim that defection increases linearly with the proportion of deceivers.

Critical strengths
  • The paper studies a relevant problem concerning the robustness of multi-agent systems against deceptive agents.

  • The authors examine different settings, providing empirical observations about how adversarial influence varies with group size, attacker proportion, and model choice.

Critical weaknesses
  • Limited novelty and experimental scope. The main experimental setting and some key conclusions have already been explored in prior studies [1,2]. The conclusion that increasing the proportion of deceivers leads to greater susceptibility is relatively intuitive, and the paper does not provide sufficiently novel insights or technical contributions beyond existing work. Furthermore, the entire study relies on a single benchmark (which is very limited) and a limited set of deliberation settings. This substantially limits the generalizability of the findings to broader multi-agent systems.

  • Insufficient evidence for the main scaling claim and its practical significance. A central claim is that adversarial influence scales linearly with the proportion of deceivers. However, Figure 3 considers only four distinct proportions, including the non-adversary baseline, leaving only three nonzero proportion values. While the reported linear fits are reasonable descriptions of these observations, such limited coverage is insufficient to establish a general linear scaling relationship. Additionally, the primary metric is the defection rate of initially correct individual agents rather than the accuracy of the final group decision. It remains unclear whether the observed trends translate into meaningful degradation in overall system performance, which is arguably the more important concern for multi-agent systems.

[1] When collaboration fails: persuasion driven adversarial influence in multi agent large language model debate. [2] On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents.

Questions for authors

I leave the other two weaknesses and some questions here.

  • The finding that private coordination makes deceivers less effective is potentially interesting, but the paper provides limited analysis of why this happens. More generally, the study is largely descriptive and offers limited mechanistic understanding of its observations.

  • The paper could be more clearly presented. In particular, Figures 1 and 2 are unnecessarily elaborate for the relatively simple experimental setup and provide limited technical insight. Figure 6 takes up a large portion of the page, with rather limited information.

  • Would this linear relationship persist across other datasets and a wider range of adversarial proportions?

  • How does the adversarial proportion affect final group accuracy or attack success rate? Does the same scaling trend hold when considering collective decisions rather than individual agent defection?

7 comments
AuthorsOpenReview ↗

Response to reviewer

We thank the reviewer for their feedback! We now turn to the weaknesses and questions raised.

W1: Relation to [1] and [2]

Unlike prior work, our study directly examines the scalability and robustness of multi-agent systems under adversarial influence. Huang et al. [2] introduce one faulty agent at a time in groups of two to five, while Kraidia et al. [1] study a single adversarial agent. Neither systematically examines how deception scales with group composition or whether adding honest agents improves robustness. Our results show that deceiver proportion predicts defection better than absolute count, meaning that scaling up a multi-agent system does not reliably improve robustness when adversaries scale alongside it. Beyond this central finding, we show that deceptive minorities degrade collective accuracy, private coordination unexpectedly reduces deception effectiveness, and model choice substantially affects susceptibility. These findings have direct implications for how multi-agent systems should be scaled, evaluated, and defended against adversarial participants.

W1: Is the result intuitive?

We agree that more deceivers causing greater harm is intuitive. However, whether that harm depends on the number of deceivers or their proportion is far less obvious, and it determines the cost and scalability of multi-agent systems. If harm depends on count, adding honest agents dilutes adversarial influence, so larger groups are a viable defense at extra inference cost. If it depends on proportion, expanding the group buys no protection when adversaries scale alongside it. Our results support the latter: proportion remains significant after controlling for count (p < 0.001), while count adds nothing (p = 0.57). Spending more compute on additional agents does not buy robustness.

W1/Q3: Single benchmark

We chose HLE because its verifiable answers let us measure precisely when deceivers mislead initially correct agents. This is also a setting where honest agents can defend their answers with evidence or reasoning. If substantial influence occurs even here, it raises broader concerns for open-ended settings such as planning and safety assessments, where correct judgments are harder to verify and are thus more susceptible to shifting. Our findings are also consistent across four model families, two deliberation protocols, and heterogeneous model pairings. Our experiments already cost over $14,000, and replicating the full grid on another benchmark would cost $20,000.

W2/Q3: Number of proportions

While four tested proportions do not establish a universally linear relationship, our claim of approximate linearity is limited to the tested range and supported by additional analyses: allowing a separate effect for each proportion did not significantly improve fit (p=0.30). More importantly, substantial influence emerges well before deceivers form a majority: even a 1/5 share increased defection by 8.1 to 14.3 points. We focus on honest majorities because at a share of 1/2 or more, deceivers block a correct majority without persuasion.

W2/Q4: Collective decisions

Final majority-vote accuracy falls as deceiver proportion increases, even though honest agents always form the majority:

Model01/51/33/7p
Gemini64.561.752.646.4<0.001
Grok53.750.043.237.2<0.001
DeepSeek52.350.546.839.3<0.001
MuseGlimmer44.136.332.122.2<0.001

Deception also erases the benefit of deliberation: without deceivers, discussion improved honest accuracy by 7 to 18 points, while at 3/7 the gain nearly disappeared for Gemini and Grok (+0.2 each) and reversed for Muse Glimmer (-11.0). Deceiver proportion also predicts these group outcomes better than count. Deception compromises the honest majority's collective judgment, not just individual answers.

Q1: Why does coordination lower defection?

Our independent-deceiver condition already includes private planning. Deceivers in both conditions receive the reference answer and privately plan every round (Appendix B.2); coordination only adds a private channel. Coordination reduced defection despite using slightly more inference (+16% tokens), and honest agents expressed suspicion of collusion in only 1 of 60 coded trials. Coordinated deceivers converged on a single wrong answer, whereas independent deceivers advanced several. Among honest agents who switched to a wrong answer, the share adopting a deceiver's answer fell from 71.6% to 49.5% under coordination (Grok). A single answer gives honest agents one target to challenge.

Q2: Figures

In the revision, we remove Figure 2 and simplify Figures 1 and 6 to make room for the new results.

Our central finding remains: an honest majority does not guarantee robust collective decisions: adversarial influence depends on the proportion of deceivers rather than number.

Sincerely, The Authors

AuthorsOpenReview ↗

Additional experiments for generalization

We thank the reviewer again for their feedback! To directly address Q3 concerning generalizability beyond HLE, we conducted new experiments across four additional benchmarks, including two safety-related benchmarks. These experiments test whether the relationship between deceiver proportion and defection persists across different tasks, answer formats, and safety-relevant settings.

Q3: Other datasets

We selected four benchmarks that differ from HLE in task type, answer format, and safety relevance:

  • SimpleQA Verified: Short-answer factual recall without tools, testing whether deception remains effective without extended reasoning.
  • SuperGPQA-hard: Ten-option multiple-choice questions spanning 285 graduate disciplines, requiring deceivers to argue for existing incorrect options rather than invent answers.
  • WMDP: Biosecurity and cybersecurity questions, extending the evaluation to hazardous-knowledge domains.
  • SALAD-Bench: Safety judgments across six harm domains, requiring agents to distinguish safe from unsafe responses to harmful requests.

Together with HLE, these benchmarks cover factual recall, complex reasoning, multiple-choice decisions, and safety judgments.

We evaluated Gemini 3.8 Flash across six group compositions: 8 honest agents with 0, 2, 4, or 6 deceivers (varying share), plus 4+3 and 12+9 at a fixed 3/7 deceiver share.

Across all four additional benchmarks, defection increased significantly with deceiver proportion (within-question permutation tests, all p<0.05), following approximately linear trends over the tested range (R2=0.86–0.95) with no significant departure from linearity (p = 0.15). The proportion model fit better than the count model on every benchmark and across all four benchmarks combined (ΔAIC = 21.3), while adding absolute deceiver count provided no significant improvement (p=0.32).

These experiments directly address the reviewer's concern about generalizability beyond HLE. Across five benchmarks, deceiver proportion consistently predicts adversarial influence better than absolute count, while increasing group size does not reliably protect an honest majority. We will incorporate these experiments into the appendix.

Our central contribution is not simply that more deceivers cause greater harm, but that adversarial influence tracks their proportion rather than absolute count, with important implications for the robustness and scalability of multi-agent systems.

We thank the reviewer for encouraging us to test our findings beyond HLE. The new benchmarks have made our central result considerably more robust, and the safety benchmarks make its real-world consequences far more salient: a deceptive minority can sway a group's judgments about hazardous knowledge and harmful requests.

Together with our earlier response, we hope these additional experiments address the reviewer's principal concerns, and we would be grateful if the reviewer would reconsider their assessment.

Sincerely,

The Authors

Thank you for the detailed rebuttal and manuscript revisions. I appreciate your efforts to address my concerns. However, I still have several concerns:

  • I understand your distinction between the absolute number and proportion of deceivers. However, I remain unconvinced that this finding provides sufficiently novel or actionable insights. Intuitively, a higher proportion of malicious agents would be expected to increase adversarial influence, while increasing the number of honest agents relative to malicious ones should improve robustness. The authors argue that adding agents does not help when adversaries scale proportionally, but this implication is also relatively straightforward. It remains unclear what non-obvious design decisions or practical benefits follow from the proposed relationship beyond these intuitions.

  • I still find the emphasis on linear scaling unconvincing. The experiments evaluate only four distinct proportions, including the zero-adversary baseline, leaving just three nonzero points. Even the additional benchmark results do not consistently exhibit a linear increase (e.g., SimpleQA Verified). Although the authors report reasonable linear fits and no significant evidence of nonlinearity, this does not establish a general linear scaling relationship. A scaling claim should be supported by substantially more extensive measurements across a wider range of proportions. I believe the current evidence supports a positive association between adversarial proportion and defection, but not a sufficiently established linear scaling law.

  • The additional majority-vote accuracy results are useful. However, if the final votes include deceivers, part of the accuracy degradation may result directly from their incorrect votes rather than from persuading honest agents. Reporting honest-only majority accuracy would help distinguish these effects.

  • Minor: I appreciate the four additional benchmarks, which substantially address my original concern about relying on HLE alone. I would also encourage you to present these results more prominently in the main paper rather than only in the appendix.

Overall, the rebuttal improves the experimental coverage and addresses some of my original concerns. Nevertheless, my main concern remains the significance of the central contribution. The paper primarily establishes that a greater proportion of malicious agents is associated with greater adversarial influence, while the stronger claim of linear scaling is not sufficiently supported. I do not find the resulting insights sufficiently novel or substantial for acceptance. Therefore, I will maintain my original rating.

Reviewer_6v8ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how deceptive agents affect multi-agent deliberation on HLE. It tests four models and twelve group compositions, ranging from 2 to 21 agents, and measures how often honest agents move from a correct initial answer to an incorrect final answer. The results show that defection increases with the proportion of deceivers, which predicts outcomes better than their absolute count in the fitted models. Additional experiments on two models show that private coordination reduces defection. There are several other lines of evaluations that look at various baselines such as all-honest baselines, matched-question comparisons, heterogeneous model pairings, and transcript analysis.

Critical strengths

The core question this paper is asking is interesting. Specifically, testing whether the number of deceptive agents or their share of the group better predicts defection. As is the the main finding of the paper that larger groups do not consistently protect correct judgments when the proportion of deceivers stays fixed, even when honest agents remain the majority.

The main methodology used to answer this question also seems valid. The authors compare different group sizes while keeping that share fixed, and run the analysis across four models and groups of up to 21 agents.

The experimental results further show that private coordination reduces defection by about 5% to 7% in matched comparisons for both tested models. This is also an interesting finding as it suggests that testing only coordinated attackers could make a system look more robust than it is against independent attackers. Although it is still unclear what causes the difference.

Critical weaknesses

The claim in the abstract is too strong. It says that proportion, rather than group size, determines susceptibility, but both fitted models include a group-size term whose estimate and uncertainty are not reported. Four tested proportions also give limited evidence for a general linear relationship. I would also like clarification on how the binomial likelihood-ratio tests handle dependence among agents in the same discussion and repeated trials on the same questions.

Each model is tested on 95 to 107 HLE questions, after removing questions answered correctly zero or four times during screening. This selects questions where responses are already variable, which limits the broader interpretation. Defection is a reasonable endpoint, but it does not tell us the net effect on performance. We also need wrong-to-correct transitions and final accuracy. The comparison with humans uses different tasks and protocols. This is acknowledged in the limitations, but the abstract still makes too much of the comparison.

Questions for authors

What are the estimated group-size effects and their confidence intervals? Does proportion still predict outcomes better than count when the analysis accounts for clustering by question and dependence within discussions?

How do wrong-to-correct transitions and final honest-agent accuracy vary across the scaling and coordination conditions? Do they support the same conclusions as the defection results?

Would the coordination effect persist against a planning-only control with a comparable inference budget? It is not clear what is driving the reduction.

8 comments
AuthorsOpenReview ↗

Response to reviewer

We thank the reviewer for their careful reading and constructive feedback!

W1/Q1: Group-size estimates

Group-size coefficients and 95% confidence intervals:

CoefficientEstimate [95% CI]
β(P)1.83 [1.34, 2.33]
β(C)0.09 [0.06, 0.13]
γ(P)0.19 [0.12, 0.27]
γ(C)0.09 [-0.01, 0.20]

Deceiver proportion predicts defection better than absolute count, even after controlling for group size. The proportion model fits substantially better (deviance difference = 61.5). Adding proportion to the count model improves fit (p < 0.001), whereas adding count to the proportion model does not (p = 0.44). Group size is positively associated with defection controlling for proportion (γ(P) = 0.19 [0.12, 0.27]), but not significantly so controlling for count (γ(C) = 0.09 [-0.01, 0.20]). Thus, larger groups do not provide protection against defection.

W1/Q1: Dependence within discussions and across questions

Deceiver proportion remains a stronger predictor after accounting for dependence. We verified this three ways:

  • Question-clustered standard errors: in a combined model, proportion remained significant (β = 1.98 [1.28, 2.68], p < 0.001) and count did not (p = 0.57).
  • Question-level bootstrap: resampling entire questions 2,000 times, the proportion model fit better in all 2,000 resamples.
  • Generalized estimating equations: with exchangeable correlation within questions, proportion again fit better (QIC 17,478 vs. 17,578; lower is better).

W1: Four tested proportions

The approximately linear relationship holds across the tested range. Allowing separate effects for each proportion did not significantly improve fit over a linear trend (pooled p = 0.30), and the trend remained significant excluding the all-honest baseline (p < 0.001). While four proportions cannot establish linearity beyond the tested range, defection increased across tested proportions, even with deceivers remaining a minority.

W2/Q2: Wrong-to-correct transitions and final accuracy

Deceivers undermine accuracy gains from deliberation, not just initially correct judgments. As deceiver proportion increased from 0 to 3/7, net accuracy gains fell across all four models, nearly disappearing for Gemini (+18.2 to +0.2 points) and Grok (+14.3 to +0.2), and becoming negative for Muse Glimmer (+7.2 to −11.0). Final accuracy declined significantly across all four models (trend p ≤ 0.05). Private coordination mitigated this decline relative to independent deception (Gemini +5.6 points [2.1, 8.8]; Grok +4.2 [0.7, 7.6]).

ModelAccuracy gain, 0 → 3/7Wrong-to-correct, 0 → 3/7Final accuracy, 0 → 3/7
Gemini+18.2 → +0.248.2% → 34.5%67.0% → 48.9%
Grok+14.3 → +0.234.9% → 25.8%56.1% → 42.5%
DeepSeek+14.9 → +8.042.3% → 39.8%55.2% → 49.8%
Muse Glimmer+7.2 → -11.028.9% → 16.8%46.8% → 29.6%

W2: Question selection

Our difficulty-stratified sampling examines susceptibility across different levels of initial reliability. The relationship between deceiver proportion and defection holds across all three difficulty buckets (5.5, 4.8, and 3.6 percentage points per 0.1 increase in deceiver proportion; all p < 0.001), including questions answered correctly in 3/4 screening attempts. Susceptibility is not confined to difficult questions, but persists across the full range of difficulty levels evaluated.

W2: Human comparison

We agree that human studies use different tasks, and this makes our findings more concerning. Those studies examined eyewitness judgments and behavioral imitation, while our agents answer questions with verifiable answers that a correct agent can defend with evidence or reasoning. Even so, a 1/5 deceiver proportion increased defection by 8.1 to 14.3 points for Gemini, Grok, and Muse Glimmer, whereas the cited human studies found no significant influence from a single misleading confederate. This raises concerns for subjective or ambiguous settings, where a correct judgment is harder to defend.

Q3: Planning-only control

Our experiment compares independent and coordinated deception as two distinct social configurations. Both allow private planning; coordination additionally permits strategy exchange, as in real-world deliberation. Matching exchanges with extra isolated reflection would test a different intervention rather than enabling communication. Reasoning-token usage per response is comparable: coordinated deceivers use 359 versus 397 tokens per planning step and 308 versus 340 per public message; honest agents use 424 versus 420. Thus, reduced deception effectiveness is not accompanied by increased reasoning effort per response.

Our central finding remains: deceiver proportion predicts defection better than count, and increasing group size does not provide reliable protection.

Sincerely,

The Authors

AuthorsOpenReview ↗

Follow-up to address comparable inference budget experiment for coordination

We thank the reviewer again for suggesting a planning-only control! As clarified in our previous response, both independent and coordinated deceivers already plan privately each round, with coordination additionally allowing strategy exchange. To distinguish the effects of communication from additional inference, we have now run the requested planning-only control.

Q3: Planning-only control

We gave independent deceivers an additional private planning pass each round, allowing them to revise their strategies without communicating with other deceivers. This matched the coordinated condition's number of model calls (3 deceiver calls per round in both conditions, versus 2 for the original independent condition).

Additional private planning did not reproduce the reduction in defection observed under coordination. Relative to independent deceivers, planning-only changed defection by just −0.8 percentage points across compositions (p=0.80, question-level resampling), with no consistent reduction. In contrast, coordination consistently reduced defection across the tested compositions despite a comparable number of model calls and, broadly speaking, inference budget.

Together with our earlier finding that coordinated deceivers use comparable reasoning tokens per response, these results support the interpretation that communication among deceivers, rather than additional inference alone, contributes to the reduction in deception effectiveness. We will incorporate the planning-only control into the appendix.

We are grateful to the reviewer for suggesting the planning-only control. It has sharpened our interpretation of the coordination results.

We hope this additional experiment, together with our earlier analyses, addresses the reviewer's remaining concerns, and we would be grateful if the reviewer would consider raising their score.

Sincerely,

The Authors

Dear Reviewer 6v8R,

Thank you again for the thoughtful questions, which prompted several useful additions. We reported the group-size estimates and checked that the scaling result holds when accounting for dependence within questions and differences in question difficulty. We also ran the planning-only control you suggested; extra private planning did not reproduce the reduction in defection seen under coordination. We hope these results resolve the methodological concerns you raised, and we'd be grateful if you would consider raising your score.

Sincerely,

The Authors

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. We summarize the central finding and new analyses added during rebuttal.

Our paper asks whether increasing group size protects language-model agents against deception. Across the tested settings, deceiver proportion predicts susceptibility better than absolute count, and deceptive minorities compromise both individual judgments and honest-majority decisions.

Our rebuttal adds five lines of evidence:

  • Scaling robustness: proportion outperforms count after controlling for group size and accounting for question-level dependence; adding more honest agents does not reliably reduce defection.
  • Collective accuracy: deceptive minorities reduce deliberation gains and lower the accuracy of honest agents' majority vote.
  • Prompt sensitivity: the proportion effect persists across calm and emotional instructions for both honest agents and deceivers.
  • Cross-benchmark tests: the relationship persists on factual recall, graduate-level knowledge, hazardous knowledge, and safety judgments.
  • Coordination control: matching the number of model calls through additional private planning does not reproduce the reduction observed with attacker communication.

These findings extend the scaling result beyond the original benchmark and prompts while clarifying its consequences for collective accuracy and system design. The new analyses are included in the revised appendix.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.