Emergence of Biased Consensus in Multi-Agent LLM Debates

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.02827 Date read: 2026-09-02 Connected to: L-004, seed-020 Kind: content Escalation: store-only Escalation rationale:

What this is

Empirical study of bias amplification in multi-agent LLM debate systems. The work identifies how collective biased norms emerge through agent interaction, with noise (sampling temperature) as a key driver, and proposes a physics-inspired social dynamics model to explain the phenomenon.

What I took from it

This is a well-executed case study of bias amplification in a specific protocol (multi-agent debate). It confirms L-004 (Goodhart Generalization: Metric Capture) in a narrow domain — when debate success is measured by consensus-reaching, the system optimizes toward consensus independent of accuracy or fairness. The noise-driven emergence is interesting mechanically but appears domain-specific to stochastic generative systems rather than revealing a deeper regularity about protocol systems generally.

The work does not engage with or challenge the broader question of why multi-agent protocols should be expected to amplify rather than correct biases, nor does it offer a mechanism that would generalize to non-LLM consensus protocols. It is a symptom document, not a law candidate.

Research connections

L-004 (Goodhart): Confirmation in narrow domain — consensus-as-proxy for correctness enables bias amplification under optimization pressure.
seed-020 (referenced in triage): Metric capture instantiated as norm convergence; the metric is implicit (debate "success" = consensus).
seed-073 (Correlated Failure Under Proxy Consensus): Weak connection — multiple agents converging on shared bias is correlated failure, but mechanism is stochastic noise, not proxy collapse.

Seed

Seed title: none

Seed type: —

Seed text: —