Conserved Verdicts, Drifting Commitments: Identifying Harmful Overthinking
Summary
- A research paper investigates harmful overthinking in reasoning models, where models that reach a correct answer lose it by continuing to think.
- The authors introduce a mathematical framework comparing a model's current commitment to its final verdict to measure expected remaining changes in answer quality.
- Testing across six reasoning models from three families on MATH-500 and OlympiadBench shows that declining accuracy curves stem from which chains remain active, while ongoing chains maintain or improve.
Why it matters
Understanding and identifying harmful overthinking in reasoning models provides a framework for developing better stopping rules to optimize inference budgets and prevent accuracy degradation.