Think Less, Locally: Local Preference Learning for Reasoning Compression

Researchers proposed LoCo, a local preference learning method for reasoning compression that detects overthinking states from a model's own traces, freezes reasoning prefixes, and trains on state-conditioned preference pairs using DPO. The approach teaches models when to continue reasoning and when to conclude, rather than simply encouraging shorter trajectories overall. Training data was constructed entirely from self-generated mathematical traces and evaluated on five mathematical and three code generation benchmarks.