Emergence of Biased Consensus in Multi-Agent LLM Debates

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.

Large language models are increasingly used not only as individual systems, but also as groups of AI agents that discuss a question before making a decision. These debates can improve performance, but they may also introduce a hidden risk: small biases from individual models can become stronger when agents influence one another. We study this problem in multi-agent LLM debates and show that biased consensus can emerge in realistic settings such as investment recommendations and AI-based judging. To explain why this happens, we build a mathematical model inspired by how physicists study collective behavior, where many small interactions can suddenly produce large-scale agreement. The model predicts when a group of LLM agents is likely to lock into a biased decision, depending on factors such as randomness, conformity, and agent diversity. Our experiments confirm these predictions and show that mixing agents with different behaviors can reduce biased lock-in. This work helps researchers and practitioners better understand, predict, and mitigate safety risks in group-based AI systems.