DoubtLess
A training-free contrastive decoding method that uses the logit difference of an auxiliary small reasoning model under concise and verbose prompts to guide a large reasoning model toward generating more concise chains of thought, shortening reasoning length without loss of accuracy.
Timeline (last 90 days)
DoubtLess is a training-free contrastive decoding method that uses a small reasoning model to guide a large model to generate more concise chain-of-thought without loss of accuracy.
The core mechanism of DoubtLess is: let an auxiliary small model read the CoT already generated by the large model under two prompts, concise and verbose, respectively, then multiply the difference between the logit vectors under the two prompts by an intensity coefficient and add it to the logits of the main model's decoding at each step.
In DoubtLess, what is contrasted is not a strong model and a weak model, but the behavior of the same auxiliary small model under concise and verbose prompts.
Taking Qwen3-32B as an example, DoubtLess achieves a 31% reduction in CoT length on seven benchmarks with accuracy remaining unchanged.
DoubtLess outperforms five inference-time baseline methods on seven benchmarks.
In a self-debugging coding agent scenario, DoubtLess reduces thinking tokens per round by 20% with task performance unchanged.
DoubtLess generalizes to two other model families, likewise without loss of accuracy.
When combined with training-based compression methods, DoubtLess can further shorten the CoT of models that have already undergone conciseness training.
Ablation experiments and behavioral analysis show that the tokens saved by DoubtLess mainly come from after the first appearance of the correct answer.
The DoubtLess paper's code has been open-sourced on GitHub.
All Facts (10)
DoubtLess is a training-free contrastive decoding method that uses a small reasoning model to guide a large model to generate more concise chain-of-thought without loss of accuracy.
50%UnverifiedThe core mechanism of DoubtLess is: let an auxiliary small model read the CoT already generated by the large model under two prompts, concise and verbose, respectively, then multiply the difference between the logit vectors under the two prompts by an intensity coefficient and add it to the logits of the main model's decoding at each step.
50%UnverifiedIn DoubtLess, what is contrasted is not a strong model and a weak model, but the behavior of the same auxiliary small model under concise and verbose prompts.
50%UnverifiedTaking Qwen3-32B as an example, DoubtLess achieves a 31% reduction in CoT length on seven benchmarks with accuracy remaining unchanged.
50%UnverifiedDoubtLess outperforms five inference-time baseline methods on seven benchmarks.
50%UnverifiedIn a self-debugging coding agent scenario, DoubtLess reduces thinking tokens per round by 20% with task performance unchanged.
50%UnverifiedDoubtLess generalizes to two other model families, likewise without loss of accuracy.
50%UnverifiedWhen combined with training-based compression methods, DoubtLess can further shorten the CoT of models that have already undergone conciseness training.
50%UnverifiedAblation experiments and behavioral analysis show that the tokens saved by DoubtLess mainly come from after the first appearance of the correct answer.
50%UnverifiedThe DoubtLess paper's code has been open-sourced on GitHub.
50%