Tackling LLM Overthinking: DASH Algorithm Boosts Math Reasoning Accuracy

Chia-Hsuan Lee · hf · 2026-08-06

Reasoning language models frequently "overthink," exhibiting behaviors like hedging, approach abandonment, or self-contradiction that consume tokens without improving the answer. The research shows that even when controlling for response length, incorrect traces exhibit higher rates of unproductive self-reflection than correct ones.

To identify helpful versus harmful reflections without costly step-level annotations, the authors propose using intermediate answer commitments within reasoning traces as a cheap proxy. Building on this insight, they introduce DASH (Drift Aware advantage SHaping), which assigns segment-level credit based on whether a reasoning segment leads toward or away from the ground truth.

On competition-level math benchmarks, DASH effectively reduces overthinking behaviors, achieves more productive self-correction than baselines, and reaches the highest average accuracy (59.45% vs. 58.1% for Dr.GRPO vs. 56.95% for GRPO) where overthinking is prevalent.

Original post →

More from Research

Research channel →