LoCo: Local Preference Learning Reduces Reasoning Length by Up to 45.4% While Maintaining Accuracy

·2026.11.05 00:38

Key point

The method uses DPO on self-generated traces to teach models when to stop reasoning, improving the accuracy-efficiency trade-off across math and code benchmarks.

Details

LoCo is a local preference learning method that detects overthinking states in a model's own traces and trains on state-conditioned preference pairs using DPO.
The approach reduces completion length by 41.7%--45.4% on mathematics and 23.4%--36.3% on code generation while maintaining or improving accuracy.
LoCo achieves the best AUC across all code benchmarks and transfers effectively to code generation despite being trained exclusively on mathematical traces.
Training data is constructed entirely from self-generated mathematical traces, and the model is available as LoCo-7B on Hugging Face.

code-generation mathematical-reasoning arxiv dpo llm-efficiency reasoning-compression

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.