Agreement on model answers often stops reasoning too early
Stable Answers, Unfinished Reasoning: Why Self-Consensus Is Not a Safe Early-Exit Signal
Computation and Language
Summary
Cutting the time it takes for AI to answer questions by stopping when it repeats the same answer seems smart, but this paper shows it is risky. The authors show that just because an AI repeats the same answer several times doesn't mean it has finished thinking; it might change its mind later. They tested many ways to decide when to stop early but none were safe except a method called DEER. So relying on repeated answers alone can cut off important corrections and lead to mistakes.
What this means in practice
- •For ai inference engineers: Avoid unreliable early-stopping conditions based on answer agreement to prevent premature halting in AI reasoning pipelines.
- •For natural language processing developers: Improve AI answer accuracy by integrating more robust stopping signals than self-consensus when designing model reasoning processes.
Authors
Yunxiang Mo, Donghao Zhao, Hejia Geng
Abstract
A natural way to cut reasoning-model inference cost is to repeatedly probe a single partial trajectory for its current answer and stop once probes agree -- self-consensus. We ask whether any such rule is both safe and token-saving, and whether one can be selected once and reused. A preregistered sweep of 3,520 consensus rules, replayed on frozen trajectories from two models and three benchmarks, clears none of three acceptance gates fixed in advance; the frontier reproduces on a held-out split and on two unseen models -- while a boundary-confidence control (DEER) swept through the same pipeline clears all three. The reason lies in the signal: agreement establishes that the current answer persists under a fixed probing procedure, not that the reasoning has terminated -- a consensus-termination gap. Stopping on it commits non-terminal answers. At a rule still saving 32% of the tokens, one stop in nine fires on an answer the trajectory itself later abandons, and most of those stops cut off a correction it would otherwise have made. Widening the agreement window does not remove them: the share levels off near 7%, and by then the saving has fallen to 8%. Probe re-wording and a hand-labelled error taxonomy show the agreed answer is often a placeholder the model had not settled on. Used on its own as the stop signal, agreement fails not because it is insufficiently strict, but because it repeatedly measures the wrong object.