- 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
- The paper was inspired by the human conformity studies and experimented on LLM agents, and finally got some unexpected results..
Critical weaknesses
-
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.
-
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
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.
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
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