AI
Conserved Verdicts, Drifting Commitments: Identifying Harmful Overthinking
Publication: cs.AI updates on arXiv.org
Author: Jianghong Zhou, Hao Rong, Chen Lin
Source published: 18/11/2026, 05:00:00 UTC
Briefing created: 18/11/2026, 06:07:11 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 18/11/2026, 05:03:25 UTC.
Briefings
A new arXiv preprint argues that reasoning models can lose accuracy by continuing to think after reaching a correct answer. The study measured this 'commitment-verdict gap' on MATH-500, finding that continuing to reason causes an average loss of 3.6 points (95% CI [2.2, 5.2]) compared to stopping. This finding suggests that final correctness alone cannot predict whether a model should halt its reasoning process.
Key points
- Continuing to reason after a correct answer leads to an average accuracy loss of 3.6 points on MATH-500.
- The loss in accuracy is concentrated in a few specific problems rather than being distributed evenly.
Context
This research addresses the 'conserved verdicts, drifting commitments' problem in large language models, where models may lose accuracy by continuing to reason after reaching a correct answer. The study provides a framework for measuring the expected remaining change in answer quality and the accuracy gained by stopping, which is crucial for designing optimal stopping rules.
Limitations
The study uses MATH-500 and OlympiadBench with budgets up to 16,384 tokens. Results are limited to three reasoning model families and do not account for all possible stopping strategies or varying model architectures.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/11/2026, 06:07:11 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieora3jmxs5pmwwymfxgtn4pd45mrf2dxsxx6svonef36rbctzgv4