DoubtLess: Training-Free Contrastive Decoding for Brevity
October 12, 2026
DoubtLess is a training-free decoding method that uses a smaller reasoning model to steer a larger model toward shorter Chain-of-Thought (CoT) sequences. It works by adding the logit difference between a 'brief' and 'verbose' version of the small model to the large model's logits.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference costs and token usage in reasoning models without requiring additional fine-tuning.
●
researcherThis provides a new method for controlling reasoning length through contrastive decoding rather than heuristic pruning.