Emergence of Biased Consensus in Multi-Agent LLM Debates

2026/08/03 by Maya Okawa
Computer Science · #cs.MA

paper · pdf

Accepted at the 43rd International Conference on Machine Learning (ICML 2026). 23 pages, 12 figures

arxiv created 2026/08/03 · arxiv updated 2026/08/05

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.

Citations

• Measuring and Mitigating Identity Bias in Multi-Agent Debate via Anonymization
• AccessEval: Benchmarking Disability Bias in Large Language Models
• Emergent Social Dynamics of LLM Agents in the El Farol Bar Problem
• Your AI Bosses Are Still Prejudiced: The Emergence of Stereotypes in LLM-Based Multi-Agent Systems
• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?
• Persona Vectors: Monitoring and Controlling Character Traits in Language Models
• Surface Fairness, Deep Bias: A Comparative Study of Bias in Language Models
• An Empirical Study of Group Conformity in Multi-Agent Systems
• Multiple LLM Agents Debate for Equitable Cultural Alignment
• The Truth Becomes Clearer Through Debate! Multi-Agent Systems with Large Language Models Unmask Fake News
• FairEval: Evaluating Fairness in LLM-Based Recommendations with Personality Awareness
• Debate Only When Necessary: Adaptive Multiagent Collaboration for Efficient LLM Reasoning
• Understanding Bias Reinforcement in LLM Agents Debate
• Unequal Opportunities: Examining the Bias in Geographical Recommendations by Large Language Models
• Voting or Consensus? Decision-Making in Multi-Agent Debate
• Emergence of human-like polarization among large language model agents
• Large Language Models Reflect the Ideology of their Creators
• Teaching Models to Balance Resisting and Accepting Persuasion
• Emergent social conventions and collective bias in LLM populations
• Mitigating the Risk of Health Inequity Exacerbated by Large Language Models
• Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge
• Towards Implicit Bias Detection and Mitigation in Multi-Agent LLM Interactions
• AI generates covertly racist decisions about people based on their dialect
• FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making
• Improving Multi-Agent Debate with Sparse Communication Topology
• Standard Language Ideology in AI-Generated Language
• The Silicon Ceiling: Auditing GPT’s Race and Gender Biases in Hiring
• Large Language Models are Geographically Biased
• The Political Preferences of LLMs
• Cultural bias and cultural alignment of large language models
• Simulating Opinion Dynamics with Networks of LLM-based Agents
• Confronting LLMs with Traditional ML: Rethinking the Fairness of Large Language Models in Tabular Classifications
• SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
• Evaluating Large Language Models at Evaluating Instruction Following
• Benchmarking Cognitive Biases in Large Language Models as Evaluators
• ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs
• AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors
• Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
• Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
• Large Language Models are not Fair Evaluators
• Improving Factuality and Reasoning in Language Models through Multiagent Debate
• Whose Opinions Do Language Models Reflect?
• The Spontaneous Emergence of Conventions: An Experimental Study of Cultural Evolution
• Statistical physics of social dynamics
• Phase transitions in social impact models of opinion formation
• Lectures on Phase Transitions and the Renormalization Group
• Dynamic models of segregation†