DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning
arXiv cs.AI · 14 Oct 2026, 03:59 AEST
DoubtLess is a training-free decoding method designed to enhance reasoning efficiency in models. It successfully shortens the chain of thought by 31% without sacrificing accuracy across various benchmarks.
Key facts
- Reasoning models achieve strong performance through long chains of thought (CoT), but often spend more tokens than needed to solve a problem.
- We propose DoubtLess, a training-free decoding method in which a much smaller reasoning model steers the large one towards brevity.
- On seven benchmarks of competition mathematics, graduate-level science and competitive programming, DoubtLess shortens the CoT of Qwen3-32B by 31% with unchanged accuracy, outperforming five inference-time baselines.
- DoubtLess reduces thinking tokens per turn by 20% in a self-debugging coding agent with unchanged task performance.
- Ablations and behavioral analysis show that DoubtLess reduces repeated checking, with most token savings coming after the correct answer first appears in the CoT.
Read aloud
DoubtLess is a training-free decoding method designed to enhance reasoning efficiency in models. It successfully shortens the chain of thought by thirty one percent without sacrificing accuracy across various benchmarks. Reasoning models achieve strong performance through long chains of thought CoT, but often spend more tokens than needed to solve a problem. We propose DoubtLess, a training-free decoding method in which a much smaller reasoning model steers the large one towards brevity. On seven benchmarks of competition mathematics, graduate-level science and competitive programming, DoubtLess shortens the CoT of Qwen3-32B by thirty one percent with unchanged accuracy, outperforming five inference-time baselines. DoubtLess reduces thinking tokens per turn by twenty percent in a self-debugging coding agent with unchanged task performance. Ablations and behavioral analysis show that DoubtLess reduces repeated checking, with most token savings coming after the correct answer first appears in the CoT.
Sources in this cluster
- DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning
arXiv cs.AI · 14 Oct 2026, 03:59 AEST