DoubtLess cuts Qwen3-32B chain-of-thought by 31% training-free with no accuracy loss

Minghao Luo · hf · 2026-10-14

DoubtLess (CUHK & EPFL) is a training-free decoding method for shorter chains of thought:

Mechanism: a small reasoner reads the large model's CoT twice — once prompted brief, once verbose — and the strength-scaled logit difference is added to the large model's logits at every step
Results: shortens Qwen3-32B's CoT by 31% with unchanged accuracy across seven math/science/coding benchmarks, beating five inference-time baselines; cuts per-turn thinking tokens by 20% in a self-debugging coding agent; generalizes to other model families and stacks with training-based compression
Analysis: savings come mostly from reduced repeated checking after the first correct answer; code is open-sourced

More from coding & agent