Know When to Stop: Segment-Level Credit Assignment for Reducing Overthinking
DASH uses intermediate answer commitments as a cheap signal for when a reasoning trace is helping or drifting.
The paper targets “overthinking” in reasoning models: hedging, abandoning approaches, and contradicting themselves while spending more tokens. Its method, DASH, assigns credit at the segment level by checking whether later reasoning moves answer candidates toward or away from the ground truth. On competition-level math benchmarks, the authors report 59.45% average accuracy, ahead of the listed baselines, while reducing unproductive reflection. HF Daily Papers' note
The paper targets “overthinking” in reasoning models: hedging, abandoning approaches, and contradicting themselves while spending more tokens. Its method, DASH, assigns credit at the segment level by checking whether later reasoning moves answer candidates toward or away from the ground truth. On competition-level math benchmarks, the authors report 59.45% average accuracy, ahead of the listed baselines, while reducing unproductive reflection. HF Daily Papers' note
score 5