Research map / Multi-agent systems and games / LLM agents

Emergence of Biased Consensus in Multi-Agent LLM Debates

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Models emergent biased consensus in multi-agent LLM debates as a finite-size crossover of a spin-model phase transition, validated with controlled logit-bias experiments and mitigated by heterogeneous sampling temperatures.

Problem

Multi-agent LLM debates can amplify individual LLM biases into collective biased norms, but the mechanisms and conditions for this emergence are poorly understood.

Approach

Reformulates Ising/Potts spin models from social dynamics for LLM debates, with conformity, local bias and sampling temperature as noise; predicts a phase transition at lambda*z/T ~ 1 that is rounded into a crossover for finite N, and shows agent heterogeneity in temperature smooths the transition.

Key results

• Low sampling temperature (T<0.5) induces early biased consensus within one or two rounds in investment and LLM-as-a-judge debates, and reduces performance.
• Heterogeneous sampling temperatures (N=6 with T in {1.4,1.5,1.6}; N=10 with T in {1.3..1.7}) reduce bias norm and improve performance relative to homogeneous-temperature baselines in real-world tasks.
• Mixing LLM types (heterogeneous populations) yields final consensus m(R) in [-0.5,0.2] versus [-0.7,0.5] for homogeneous populations at T=1.2.

Surprising findings

• Mixing LLM types or temperatures weakens biased consensus in a way predicted by smoothing of the mean-field response, rather than by any explicit debiasing.

Limitations

• Simplified debate protocol: short-term memory, all-to-all or random interactions, no explicit expert roles, discretized outputs to binary/multi-choice states.
• Only two synthetic and two realistic tasks evaluated; higher-stakes domains not tested.
• Investment evaluation window is short (about five weeks), making returns sensitive to transitory market movements.

Experimental setup

Datasets: MT-Bench human judgments implicit gender bias dataset (task sets to names) investment recommendation task prompts

Benchmarks: MT-Bench LLM-as-a-judge investment recommendation (market-adjusted return) Binary Choice synthetic task Implicit Bias task

Metrics: absolute collective norm |m(R)| self bias US/technology concentration market-adjusted return accuracy vs human labels

Baselines: homogeneous-temperature debate

Builds on: multi-agent LLM debate Ising model Potts model mean-field approximation

Closest papers by meaning

• Multi-Agent Debate for LLM Judges with Adaptive Stability Detection (NeurIPS 2025 · similarity 0.88)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.88)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.88)
• Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate (NeurIPS 2026 · similarity 0.87)
• The Value of Variance: Mitigating Debate Collapse in Multi-Agent Systems via Uncertainty-Driven Policy Optimization (ICML 2026 · similarity 0.87)
• Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.87)