PaperlayerView exact revision on arXiv

Source-linked AI summary

Do as We Do, Not as You Think: the Conformity of Large Language Models

Zhiyuan Weng, Guikun Chen, Wenguan Wang

arXiv:2501.13381v2cs.CL

TL;DR

Conformity in LLM-driven multi-agent systems is a relatively unexplored question with implications for collective problem-solving and ethical decisions. The paper introduces BENCHFORM to study conformity, its influences, and potential mitigations, and finds that conformity can affect system performance even when agents know the correct answers.

  • Problem

    A comprehensive investigation of conformity in LLM-driven multi-agent environments is absent, despite concerns that group influence may affect collective problem-solving and ethical decisions.

    Source: §1 · pp. 1–2

  • Method

    The paper introduces BENCHFORM, a reasoning-intensive benchmark with five interaction protocols, and evaluates representative LLMs while studying conformity factors and preliminary mitigation strategies.

    Source: §1 · p. 2

  • Results

    Conformity exists in LLM multi-agent systems and can substantially affect performance; models may adopt majority opinions despite knowing correct answers.

    Source: §8 · p. 10, §7 · p. 10

  • Takeaways & Limitations

    The findings motivate strategies that encourage consistency between an LLM’s initial response and its post-interaction explanation to support independent decision-making.

    Source: §7 · p. 10

    BENCHFORM and its protocols provide a necessary but not sufficient test, and may not capture varied real-world interactions, including cases where agents answer before seeing others’ responses.

    Source: §7 · p. 10

Abstract

from arXiv · show

Recent advancements in large language models (LLMs) revolutionize the field of intelligent agents, enabling collaborative multi-agent systems capable of tackling complex problems across various domains. However, the potential of conformity within these systems, analogous to phenomena like conformity bias and groupthink in human group dynamics, remains largely unexplored, raising concerns about their collective problem-solving capabilities and possible ethical implications. This paper presents a comprehensive study on conformity in LLM-driven multi-agent systems, focusing on three aspects: the existence of conformity, the factors influencing conformity, and potential mitigation strategies. In particular, we introduce BenchForm, a new conformity-oriented benchmark, featuring reasoning-intensive tasks and five distinct interaction protocols designed to probe LLMs' behavior in collaborative scenarios. Several representative LLMs are evaluated on BenchForm, using metrics such as conformity rate and independence rate to quantify conformity's impact. Our analysis delves into factors influencing conformity, including interaction time and majority size, and examines how the subject agent rationalizes its conforming behavior. Furthermore, we explore two strategies to mitigate conformity effects, i.e., developing enhanced personas and implementing a reflection mechanism. Several interesting findings regarding LLMs' conformity are derived from empirical results and case studies. We hope that these insights can pave the way for more robust and ethically-aligned collaborative AI systems. Our benchmark and code are available at BenchForm.

1 INTRODUCTION

Conformity in LLM multi-agent systems is a largely unexplored concern because it may affect collective judgments on consequential social issues. The paper studies whether conformity occurs, what influences it, and how it might be mitigated.

1 INTRODUCTION

  • Conformity may support consensus and coordination, but can also produce herding that compromises judgments on voting, policy recommendations, or ethical decisions.The paper frames these risks in the context of multi-agent systems considered for public policy, moderation, and governance roles.Source: §1 · p. 1
  • Prior work has noted conformity in some scenarios, but a comprehensive investigation of the phenomenon in LLM-driven multi-agent environments is absent.The motivating question is whether multi-agent systems encounter problems that a single agent would not.Source: §1 · pp. 1–2
  • BENCHFORM evaluates reasoning-intensive tasks across five interaction protocols, using accuracy, conformity rate, and independence rate to study conformity and its influencing factors.The study also examines mitigation strategies, including enhanced personas and reflection.Source: §1 · p. 2

2 BENCHFORM

BENCHFORM is a reasoning-intensive benchmark for studying conformity in multi-agent collaboration. It combines varied tasks with protocols that model both immediate peer influence and relationships formed over multiple rounds.

2 BENCHFORM

  • BENCHFORM evaluates conformity in collaborative environments through reasoning tasks and protocols designed to probe agent relationships across short- and long-term interactions.The benchmark is derived from BBH and is intended to examine conformity-related behavior.Source: §2 · p. 2

2.1 DATA COLLECTION

  • The dataset contains 3,299 multiple-choice questions across logical and analytical reasoning and language and contextual understanding, with up to 300 samples per task type.The two categories contrast clear logically derived answers with tasks containing subjective elements.Source: §2.1 · pp. 2–3

2.2 PROTOCOLS

  • Raw Protocol establishes a no-peer baseline; Correct and Wrong Guidance test immediate responses to peers’ correct or incorrect answers, while Trust and Doubt test responses after peers’ accuracy or inaccuracy histories.Trust presents initially correct peer answers followed by an incorrect answer; Doubt reverses that sequence.Source: §2.2 · p. 3

2.3 IMPLEMENTATION DETAILS

  • The implementation positions the subject agent to respond after peer agents, drawing inspiration from Asch conformity experiments.The Raw protocol uses no additional agents; Correct and Wrong Guidance add six, while the longer protocols add historical discussions.Source: §2.3 · p. 4

3 CONFORMITY ON BENCHFORM

The study evaluates 11 LLMs on BENCHFORM using accuracy, conformity, and independence metrics. All evaluated models show conformity, with results varying by protocol and model size.

3.1 TARGET LLMS

  • The evaluation included 11 popular LLMs: GPT-3.5 and GPT-4o, plus open-source Llama3, Llama3.1, Gemma2, and Qwen2 series.The largest version of each LLM series was selected as its representative for the reported findings.Source: §3.1 · p. 4, §3.3 · p. 5

3.2 EVALUATION METRICS

  • Accuracy measures performance under protocol P, conformity rate measures initially correct Raw answers that become wrong under P, and independence rate tracks correct answers across protocols.Independence rate includes only Trust and Doubt protocols.Source: §3.2 · p. 5

3.3 MAIN RESULTS AND FINDINGS

  • Across five representative LLMs, average conformity rates were 23.5% for Wrong Guidance, 31.3% for Trust, and 47.2% for Doubt; all evaluated LLMs showed a tendency to conform.Doubt induced the most errors in most cases among the three misleading protocols.Source: §3.3 · p. 5, Table 2 · p. 5
  • As model size increased, independence rates rose; Qwen2’s rate increased from 19.6% at 7B to 57.6% at 72B, while Llama3.1-405B reached 56.1%.The paper says this pattern might indicate that larger LLMs are more capable of independent decisions.Source: §3.3 · p. 5, Figure 4 · p. 5
  • Models showed distinct conformity profiles: Qwen2-7B had CRC of 98.7% and CRD of 27.5%, while Llama3.1-70B had CRW of 9.2%.The authors suggest larger, high-quality datasets and long-context pre-training may contribute to Llama3.1’s resistance to external guidance, but say the causes remain unclear.Source: §3.3 · p. 6, Table 3 · p. 6

4 WHAT IMPACTS CONFORMITY?

The section tests whether discussion rounds and majority size affect conformity, then examines how agents explain conforming and how their responses differ. Longer discussions and larger majorities generally increase conformity, though one dissenting agent can have model-dependent effects.

4.1 INTERACTION TIME

  • Llama3-70B’s conformity rose with more discussion rounds: CRT increased from 33.9% to 44.4% and CRD from 62.3% to 69.9%, while IR fell from 35.1% to 28.6%.Qwen2-72B’s IR also fell, from 61.1% to 57.6%, as rounds increased from one to five.Source: §4.1 · p. 6

4.2 PEER PRESSURE

  • Larger majorities increased conformity more than longer discussions for Llama3-70B: CRD fell from 69.9% to 32.6% as majority size fell from six to three, versus 69.9% to 62.7% as rounds fell from five to one.The shift from majority size five to six is attributed to the absence of dissenting opinions increasing peer pressure; Figure 5 presents the influencing-factor ablations.Source: §4.2 · pp. 6–7, Figure 5 · p. 7

4.3 BEHAVIORAL STUDY

  • The behavioral study sampled up to five conforming cases per task and protocol, totaling 514 examples, and asked agents to explain their answers and whether they would change them.Responses were categorized by whether agents acknowledged conformity and whether they changed their original answer.Source: §4.3 · p. 7
  • Llama3-70B acknowledged conformity and changed its answer in 50.8% of cases, while Qwen2-72B had only one such case and 222 deny-and-stick responses.Llama3-70B maintained its original answer in 76.6% of cases where it denied conformity; case studies also found it could choose a wrong answer despite correct reasoning, while Qwen2-72B linked rare stance changes to reflection.Source: §4.3 · p. 7, Table 4 · p. 7

5 WAYS TO MITIGATE CONFORMITY

The paper explores two lightweight prompt-based approaches: an independent persona before a decision, and double-checking and reflection afterward. It reports that generalizing these strategies across models and interaction scenarios remains challenging.

5 WAYS TO MITIGATE CONFORMITY

  • The paper explores two prompt-based conformity-mitigation strategies: fostering an independent persona before a decision and prompting double-checking and reflection afterward.Figure 6 reports conformity-rate and independence-rate metrics under all protocols for the double-checking and reflection strategy.Source: §5.3 · p. 8, Figure 6 · p. 8

5.3 DISCUSSION OF MITIGATION STRATEGIES

  • The authors identify generalizability across model characteristics and interaction scenarios as a challenge for the strategies.Prompts encouraging reasonable doubt in credulous Qwen2-7B or trust in inherently suspicious Llama3 models are described as difficult to design; Figure 7 reports both metrics under all protocols.Source: Figure 7 · p. 8, §5.3 · p. 9
  • Proposed future directions are training LLMs with high-quality, large-scale data and exploring alignment strategies that make LLMs aware of conformity.The paper cites the Llama-3 and Llama-3.1 comparison as empirical support for the training-data direction.Source: §5.3 · p. 9

6 RELATED WORK

The related work situates conformity as a knowledge-conflict and multi-agent collaboration problem, while noting that systematic study of conformity in LLM-driven systems remains scarce. The paper distinguishes its work by examining general problem-solving, longer interactions, causes, and mitigation.

Knowledge Parameterization and Conflicts in LLMs

  • This study examines discrepancies between other agents’ evidence and an LLM’s internal knowledge, finding representative models highly susceptible to information from other agents.The paper frames this as a further type of knowledge conflict alongside conflicts involving external evidence or user-provided context.Source: §6 · p. 9

LLM-driven Multi-agent Systems

  • LLM multi-agent systems are also used for world simulation and collaborative problem-solving, including applications in software development and social media platform design.The paper describes agents with distinct capabilities or expertise collaborating on complex tasks.Source: §6 · p. 9

Conformity in LLM-driven multi-agent systems

  • Prior work on conformity in LLM-driven multi-agent systems is scarce: three works touched on it, through scenario showcases or simulations of cross-national collaboration and debate.The paper characterizes these studies as identifying potential issues and raising awareness, without investigating underlying causes or proposing solutions.Source: §6 · p. 9
  • Earlier conformity studies focused on specific decision scenarios, whereas this work broadens the scope to general problem-solving and includes short- and long-term protocols.It also proposes conformity-oriented metrics and a structured approach covering problem identification, causal analysis, and solution development.Source: §6 · p. 9

7 DISCUSSION

The discussion presents conformity as both a route to consensus and a risk to independent judgment, and relates longer interactions to the influence of historical context. It also qualifies the study’s scope and identifies directions for broader evaluation and mitigation.

The Duality of Conformity

  • Conformity may promote consensus and cohesive outcomes, but can compromise judgment on critical social issues such as voting and policy recommendations.The paper identifies preserving independent judgment where needed while mitigating output homogenization as an important balance.Source: §7 · p. 10

Implications for Context Attention Mechanisms

  • Longer interaction effects and CR/IR results suggest that LLM responses are influenced by historical context as well as immediate input, raising questions about attention to relevant versus irrelevant history.The authors call for reevaluating context-attention mechanisms to improve selective weighting of historical context against current input.Source: §7 · p. 10

Limitations

  • The authors report that their benchmark and protocols are a necessary but insufficient test of conformity, and that multiple-choice questions may limit generalizability.Real-world interactions may not expose agents to others’ answers before they respond, and the setup may not capture varied contexts and inputs.Source: §7 · p. 10
  • The authors cannot draw definitive conclusions about why conformity varies across architectures because specific training processes are not publicly available.They call for community collaboration to study conformity and provide a foundation for further research and discussion.Source: §7 · p. 10

Future Work

  • The authors propose expanding BENCHFORM beyond reasoning-intensive tasks and adding protocols that better mimic collaborative interaction, such as group reasoning and argumentation.They identify MMLU-Pro and other domains as possible additions.Source: §7 · p. 10

8 CONCLUSION

The study finds conformity in LLM-driven multi-agent systems, identifies factors associated with its degree, and reports promise in two mitigation strategies.

8 CONCLUSION

  • Conformity exists and can substantially impact the performance of LLM-driven multi-agent systems.Source: §8 · p. 10
  • Conformity varies with intrinsic model characteristics and extrinsic factors, including interaction time and majority size.Source: §8 · p. 10
  • Enhanced persona development and reflection mechanisms show promise for reducing conformity effects.Source: §8 · p. 10

ETHICS STATEMENT

The authors flag the potential misuse of BENCHFORM and distinguish its controlled research tasks from specific real-world situations.

ETHICS STATEMENT

  • BENCHFORM could be misapplied to design systems more prone to groupthink, with unintended real-world consequences.Source: ETHICS STATEMENT · p. 11
  • BENCHFORM tasks are not intended to represent specific real-world situations involving misinformation or harmful content.Source: ETHICS STATEMENT · p. 11
  • The authors urge users to follow ethical standards, including transparency, accountability, and fairness.Source: ETHICS STATEMENT · p. 11

REPRODUCIBILITY STATEMENT

The reproducibility statement identifies model checkpoints, repeated trials, and prompt documentation; the appendix outlines supporting material and background on Asch's conformity experiments.

REPRODUCIBILITY STATEMENT

  • Experiments used nine open-source LLMs from Ollama and specified versions of GPT-3.5 Turbo and GPT-4o, with three BENCHFORM runs and means and variances reported.Source: REPRODUCIBILITY STATEMENT · p. 11

SUMMARY OF THE APPENDIX

  • The appendix lists additional material on Asch experiments, BENCHFORM, experiment results, majority size, behavioral classifications, mitigation, and examples.Source: SUMMARY OF THE APPENDIX · p. 17

A THE ASCH CONFORMITY EXPERIMENTS

  • Asch's experiments found that about 37% of participants conformed to an incorrect group judgment at least once; smaller groups and dissenting opinions reduced conformity pressure.Source: A THE ASCH CONFORMITY EXPERIMENTS · p. 17

B.1 DATA

The appendix describes BENCHFORM's test data, model settings, and prompt design across five evaluation protocols.

B.1 DATA

  • BENCHFORM contains 13 tasks and uses 3,299 test examples, with five additional examples per task selected for prior discussions.Source: §B.1 · p. 17

B.2 MODEL

  • Closed-source models used temperature 0.7 and named GPT versions, while open-source models used Ollama defaults and the latest instruct-q4 0 versions.Source: §B.2 · p. 17

B.3 COMPLETE PROMPT

  • The authors designed 21 opinion expressions and provide prompts for five protocols; Trust and Doubt prompts delimit prior discussion with markers.Source: §B.3 · p. 17

C.1 FULL RESULTS

Across the reported experiments, longer discussions and larger majorities generally increased conformity, though majority-size effects varied by model. Empowered personas and reflection improved independence for several models, while behavioral responses showed that agents often denied conformity and maintained confidence in their original answers.

E SUPPLEMENTAL DETAILS ON BEHAVIORAL STUDY

  • The supplemental behavioral study classified answers by admission or denial of conformity and whether the agent changed or retained its original answer.Llama3.1-405B was used to classify Llama3-70B and Qwen2-72B responses, followed by manual sampling inspection.Source: E SUPPLEMENTAL DETAILS ON BEHAVIORAL… · p. 22

F.1 FULL PROMPTS AND RESULTS FOR EMPOWERED PERSONA

  • The selected empowered-persona prompt, P3, instructs the agent to cross-check others’ answers against its knowledge and remain open to agreement when their answers are correct.Three system prompts were presented, and P3 was selected as the final version.Source: §F.1 · p. 23, Table S7 · p. 23

F.2 FULL PROMPTS AND RESULTS FOR REFLECTION

  • The selected reflection prompt, P5, asks the agent to re-evaluate its answer using its own knowledge and prioritize independent judgment while accepting evidence-supported reasoning.Two reflection prompts were presented, and P5 was selected as the final version.Source: §F.2 · p. 23, Table S8 · p. 23

G.1 INTERACTION TIME

  • GPT-4o’s CRD increased from 17.8% after one round to 26.6% after five, while CRT changed only from 37.7% to 37.9%; Llama3.1-405B’s CRD rose from 28.3% to 43.2%.All three reported models showed increasing conformity tendencies with longer interaction time.Source: §G.1 · p. 24, Figure S5 · p. 24

G.2 PEER PRESSURE

  • When majority size rose from 3 to 6, GPT-4o’s CRT increased from 17.3% to 37.9% and its IR fell from 81.5% to 55.4%; Llama3.1-405B and GLM-4-Plus also showed conformity increases.For Llama3-70B, majority-size impact was minimal, while it was significant for Qwen2-72B; majority composition also mattered, with effects varying by model.Source: D MAJORITY SIZE ABLATION ON BENCHFORM · p. 22, §G.2 · p. 24, Figure S6 · p. 25

G.3 BEHAVIORAL STUDY

  • GPT-4o denied conformity and stuck to its original answer in 72.5% of cases, and Llama3.1-405B and GLM-4-Plus also often denied conformity while maintaining their answers.The behavioral study classified responses by whether agents admitted or denied conformity and changed or stuck to their original answer.Source: E SUPPLEMENTAL DETAILS ON BEHAVIORAL… · p. 22, §G.3 · p. 25, Table S9 · p. 25

G.4 EMPOWERED PERSONA

  • With empowered personas, GPT-4o’s CRT fell from 37.9% to 23.8% and its IR rose from 55.4% to 74.6%; Llama3.1-405B and GLM-4-Plus showed similar improvements in IR.For Llama3.1-405B, CRC and CRD fell while IR rose; GLM-4-Plus also had lower CRT and CRD and higher IR.Source: §G.4 · p. 25, Figure S7 · p. 25

G.5 DOUBLE-CHECKING AND REFLECTION

  • For GPT-4o, reflection reduced CRT from 37.9% to 11.0% and CRD from 26.6% to 11.6%, while IR rose from 55.4% to 84.6%; GLM-4-Plus showed similar changes.Llama3.1-405B improved on CRD and IR, but its other three CR measures increased.Source: §G.5 · p. 26, Figure S8 · p. 26

Related papers

Loading 2501.13381v2…