Think Less, Locally: Local Preference Learning for Reasoning Compression
LoCo cut completion length by up to 45.4% on math and 36.3% on code while maintaining or improving accuracy, the authors report.
The method identifies points in a model’s reasoning where further deliberation may be unnecessary, then generates shorter, verified continuations from those same points. Preference training teaches the model when to continue and when to conclude. Trained entirely on self-generated mathematical traces, LoCo was evaluated on five math and three code generation benchmarks, with gains extending to code despite math-only training. The authors say the code and model will be publicly released.
HF Daily Papers' note
The method identifies points in a model’s reasoning where further deliberation may be unnecessary, then generates shorter, verified continuations from those same points. Preference training teaches the model when to continue and when to conclude. Trained entirely on self-generated mathematical traces, LoCo was evaluated on five math and three code generation benchmarks, with gains extending to code despite math-only training. The authors say the code and model will be publicly released.
HF Daily Papers' note
score 5