Search papers, labs, and topics across Lattice.
The authors evaluate whether self-consensus鈥攈alting inference when repeated probes of a partial chain-of-thought yield identical intermediate answers鈥攕erves as a viable early-exit heuristic across a preregistered sweep of 3,520 rules. They find that self-consensus fails all predefined safety gates across multiple models and benchmarks, frequently committing to premature answers that the model's unprompted trajectory would have self-corrected. Widening the agreement window collapses token savings from 32% down to 8% while still leaving a ~7% error rate, proving that intermediate agreement measures transient placeholder stability rather than actual reasoning completion.
Stopping reasoning models early via intermediate answer consensus routinely truncates self-correction, because repeated agreement on a partial trajectory reflects prompt persistence rather than completed reasoning.
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.
