AP41 · Conformist debate

Family: Loop, stopping and orchestration

Strong evidence

Original names: conformity collapse, premature consensus, multi-agent debate without a stopping criterion

Agents debate to correct one another and end up agreeing on the most confident answer, even a wrong one.

Definition

Several agents are set to debate in the hope that they will correct one another. In practice they fall in line with the majority or the most confident view, and the group often has no stopping criterion.

Observable symptoms

  • Answers that were right in the first round turn wrong in later rounds.
  • Consensus reached very quickly.
  • Justifications that copy those of the peers.
  • A final result that does no better than a vote on the initial answers.

Why it hurts

Models are sycophantic and conformist: one persuasive, or compromised, agent drags the others along. Debate adds no information, it only redistributes confidence, and conformity grows with the length of the exchange and the size of the majority.

A defensible choice when…

  • An open question where the goal is to gather arguments rather than an answer, and a human makes the call.

Figures

  • Models frequently shift from correct to incorrect answers in response to their peers’ reasoning, even when strong models are in the majority (September 2025 preprint).
  • Conformity grows with the length of the interaction and the size of the majority; an assertive persona and individual reflection reduce it (ICLR 2025).
  • Across 7 benchmarks, majority voting explains most of the gain attributed to debate (NeurIPS 2025).

Remedies from the catalogue

  • B4 · Sampling and voting — Independent answers, then a vote, with no exchange at all between the agents.
  • E5 · Group chat and debate — If debate is still worth it: a blind first round, anonymised authors, a capped number of rounds, and a judge who scores the arguments rather than the majority.
  • B5 · Generator-evaluator — When an external oracle exists (tests, computation, source), it settles the question, not the consensus.

Measures outside the catalogue

  • Models from different families in the group.

Misused patterns

Related anti-patterns

Sources

Content reviewed on 2026-10-05

Anti-pattern catalogue