DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning
A small reasoning model is used at decode time to push a larger one toward shorter chains of thought without retraining.
DoubtLess compares “brief” and “verbose” continuations from the smaller model, then adds that contrast to the larger model’s logits during reasoning. On seven benchmarks, it cut Qwen3-32B’s chain-of-thought length by 31% while preserving accuracy. The authors also report a 20% reduction in thinking tokens per turn for a self-debugging coding agent, again without task-performance loss. Most savings came after the correct answer had already appeared in the chain of thought. HF Daily Papers' note
DoubtLess compares “brief” and “verbose” continuations from the smaller model, then adds that contrast to the larger model’s logits during reasoning. On seven benchmarks, it cut Qwen3-32B’s chain-of-thought length by 31% while preserving accuracy. The authors also report a 20% reduction in thinking tokens per turn for a self-debugging coding agent, again without task-performance loss. Most savings came after the correct answer had already appeared in the chain of thought. HF Daily Papers' note
score 4