Minority Sentinel: When to Overturn Majority Voting in Multi-Agent LLM Debates

Summary

States explicitly that multi-agent debate with majority voting rests on the Condorcet Jury Theorem's independent-errors assumption, and that shared pretraining corpora break it, causing the majority to systematically suppress correct minority opinions - a 'Minority Truth' effect present in roughly one in four divergent cases across six benchmarks, worth a 10-percentage-point theoretical recovery margin. Their Minority Sentinel meta-classifier recovers suppressed minorities at 81.2% flip precision with positive net gain on all six datasets. Evidence both that consensus aggregation destroys information and that traceable deliberation logs are themselves a usable epistemic artifact.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-08-29 literature sweep as a medium-priority candidate for the Architecture as Epistemology (ISR) paper (angle: aggregating LLM judgments; correlated errors / homogeneity across LLM agents). Verified against: Opened https://arxiv.org/abs/2606.29270 (arXiv abstract page, v1 28 Jun 2026)

Critique / open questions

Queued — not yet assessed.

Key quotes

Multi-Agent Debate (MAD) with Majority Voting is a dominant paradigm for improving LLM reasoning, yet its effectiveness rests on the Condorcet Jury Theorem's assumption of independent errors. (abstract)

Because contemporary LLMs share similar pretraining corpora, their errors are strongly correlated, causing the majority to systematically suppress correct minority opinions, a phenomenon we term Minority Truth. (abstract)