New paper finds reasoning models can get worse after they already have the answer
zainhas · x · 2026-07-26
A post highlights a new paper, Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models. The paper argues that reasoning models do not always benefit from longer chains of thought after they have already found the correct answer.
From the abstract shown in the image:
- The authors define a prefix-level trajectory evaluation protocol to separate harmless verbose reasoning from harmful overthinking.
- They find many benchmark cases require surprisingly little reasoning once the answer is reached.
- Stopping at the first correct prefix can improve accuracy by up to 21% over standard reasoning.
- Early stopping strategies can cut verbose overthinking by up to 50%, but still do not fully solve harmful overthinking.
- The failure analysis suggests correctness deviations come mainly from logical drift and visual reinterpretation.
- The findings generalize beyond multimodal benchmarks to language-only reasoning tasks.
The paper is from Simone Caldarella, Davide Talon, Rahaf Aljundi, Elisa Ricci, and Massimiliano Mancini.
More from Research
- AI slop is already clogging PR review and weakening the credit system behind science — rbhar90 · 2026-07-27
- ICML 2026 oral paper replication scores stay middling after a stricter re-scoring — profjamesevans · 2026-07-27
- Long-running agents will need immutable event logs, this thread argues — sebpaquet · 2026-07-27
- Seed IQ navigates Doom II, prompting questions about benchmarks beyond ARC-AGI — Fit_Transition8824 · 2026-07-27
- Agentic Data Science in Practice: Agents Write Code but Answer Wrong Questions — hugobowne · 2026-07-27
- A concise canon of foundational papers in ML, systems, NLP, speech, and audio — deliprao · 2026-07-27