ICML 2026 paper finds multi-agent LLM debates undergo a phase transition into biased consensus

Hidenori8Tanaka · x · 2026-11-20

Maya Okawa's ICML 2026 paper "Emergence of Biased Consensus in Multi-Agent LLM Debates" (arXiv:2608.02827) — inspired by reading Arendt's Eichmann in Jerusalem and translating the "banality of evil" into multi-agent dynamics — shows that interaction in multi-agent LLM debates can amplify individual model biases and produce collective biased norms, with sampling-temperature noise as a key driver. Using physics-inspired models of social dynamics, the author predicts a phase transition to collective bias once conformity passes a critical threshold set by initial bias and noise; controlled experiments observe a finite-size crossover consistent with this transition. Agent heterogeneity suppresses the emergence by smoothing the transition, and the insights generalize to realistic tasks like investment decisions and LLM-as-a-judge evaluation. 23 pages, 12 figures.

Original post →

More from AGI Musings

AGI Musings channel →