mlsanityData to 14 Oct 2026

Stable Answers, Unfinished Reasoning: Why Self-Consensus Is Not a Safe Early-Exit Signal

arXiv 2609.09989 ↗First posted 9 Sep 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 14 of its 27 references that are in the map. First posted on arXiv 9 Sep 2026.

Builds on

14 of 27 references.

References · 14

  1. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  2. 2023Let's Verify Step by Stepcited by 3,789
  3. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,986
  4. 2025s1: Simple test-time scalingcited by 1,341
  5. 2023Large Language Models Cannot Self-Correct Reasoning Yetcited by 1,017
  6. 2024Large Language Monkeys: Scaling Inference Compute with Repeated Samplingcited by 991
  7. 2023Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotationscited by 928
  8. 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMscited by 570
  9. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  10. 2023Let's Sample Step by Step: Adaptive-Consistency for Efficient Reasoning and Coding with LLMscited by 115
  11. 2024Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoningcited by 108
  12. 2026Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Modelscited by 7
2 more
  1. 2026Early Stopping for Large Reasoning Models via Confidence Dynamicscited by 6
  2. 2026When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Modelscited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Answer Convergence as a Signal for Early Stopping in Reasoning
  2. 2026LLMs as a Jury: Cross-Model Consensus Can Outperform Process Reward Models for LLM Reasoning
  3. 2026Reasoning Concentrates Errors, and Self-Consistency Never Notices
  4. 2026One Axis, No Brake: Self-Knowledge Limits the Filtering of Harmful Peer Conformity in LLMs
  5. 2026CITE: Anytime-Valid Statistical Inference in LLM Self-Consistency
  6. 2025Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs
  7. 2026Beyond Consensus: Trace-Level Synthesis in Mixture of Agents
  8. 2026ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling