DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning
A smaller reasoning model is used at decode time to push a larger one toward shorter chains of thought without retraining.
DoubtLess runs the small model in “brief” and “verbose” modes, then adds the contrast between those logits to the larger model’s logits during reasoning. The authors report a 31% CoT reduction for Qwen3-32B across seven math, science, and programming benchmarks with unchanged accuracy. In a self-debugging coding agent, it cut thinking tokens per turn by 20% without changing task performance. The paper says most savings came after the correct answer had already appeared in the chain of thought. ArXiv · AI/CL/LG's note
DoubtLess runs the small model in “brief” and “verbose” modes, then adds the contrast between those logits to the larger model’s logits during reasoning. The authors report a 31% CoT reduction for Qwen3-32B across seven math, science, and programming benchmarks with unchanged accuracy. In a self-debugging coding agent, it cut thinking tokens per turn by 20% without changing task performance. The paper says most savings came after the correct answer had already appeared in the chain of thought. ArXiv · AI/CL/LG's note
score 5