DoubtLess Uses a Small Model to Make Large Reasoners Think Less

Introduction

Reasoning models can solve difficult problems by producing long chains of thought, but length is not always the same as useful computation. A model may reach the correct answer and then continue checking, restating, or re-deriving points it has already settled. DoubtLess targets this redundant tail without requiring additional training for the target model.

How the method works

DoubtLess is a training-free contrastive decoding technique. A much smaller reasoning model observes the chain of thought generated so far by the larger model and produces two next-token distributions:

• one after being asked to think briefly;
• one after being asked to think in a more verbose and complete way.

The difference between these two logit vectors is scaled by a steering strength and added to the large model’s logits at every reasoning step. The large model is therefore nudged toward token choices associated with the brief-versus-verbose contrast, rather than being forced to stop at a fixed position or having words such as “wait” banned.

This distinction matters. Hard stopping rules can terminate useful reasoning, while fixed hidden-state steering may not adapt to the current context. DoubtLess instead applies a context-dependent signal during generation. The accompanying small model adds roughly 4% to the reported total FLOPs.

Reported results

Using Qwen3-0.6B to steer Qwen3-32B, the method removed 31% of thinking tokens across seven benchmarks covering competition mathematics, graduate-level science, and competitive programming. The reported accuracy remained unchanged, and DoubtLess outperformed five inference-time baselines.

The authors also tested a self-debugging coding agent. Thinking tokens per turn fell by 20% while task performance stayed unchanged. The approach generalized to two other model families with unchanged reported accuracy, and it produced additional compression when applied to a model that had already been trained to reason more concisely.

Behavioral analysis points to a specific source of the savings: most removed tokens appear after the correct answer has already emerged. In other words, the method primarily reduces repeated checking rather than deleting the central derivation needed to reach the answer.

Why it matters

DoubtLess offers a practical inference-time route to reasoning efficiency. Developers do not need to retrain every target model or build a model-specific compression dataset. A smaller model can act as a lightweight behavioral controller for a larger one, potentially making existing reasoning systems cheaper to serve.

The results should still be interpreted within the reported experimental scope. The benefit may depend on the choice of auxiliary model, the steering strength, the target model family, and the latency cost of running the extra predictions. Shorter chains may also require careful evaluation in settings where a complete and auditable reasoning trace is important.

The broader lesson is that efficient reasoning does not necessarily mean cutting off the model early. It can also mean identifying when the model has already reached a stable answer and reducing the redundant verification that follows.

Source: Hugging Face Daily Papers