Reasoning models often keep thinking long after they have what they need, and every extra word costs time and computing power. A preprint from Tencent’s WeChat AI team and Xi’an Jiaotong University, posted on 3 November, proposes a fix called LoCo that trims the excess one step at a time.
Instead of shortening a whole reasoning trace, LoCo searches the model’s own traces for “overthinking states”: points where the latest step shows both low confidence and high volatility, measured from the model’s own word probabilities. It keeps the reasoning up to that point, generates a shorter continuation, checks that the answer is still correct, then trains the model to prefer the shorter version using direct preference optimisation (DPO), a method that learns from pairs of better and worse answers.
Applied to a 7-billion-parameter open model, DeepSeek-R1-Distill-Qwen-7B, LoCo cut answer length by 41.7% to 45.4% on five maths benchmarks and by 23.4% to 36.3% on three coding benchmarks. Accuracy held roughly level or rose: on the AIME 2024 maths contest it went from 50.8% to 54.2%.
Why it matters: shorter reasoning means cheaper, faster answers. The results come from one small model, are the authors’ own and have not been peer reviewed; code is promised but not yet released. For why models reason out loud at all, see our Chain-of-Thought Prompting paper card.
Source: Yuan et al., Think Less, Locally: Local Preference Learning for Reasoning Compression (arXiv)