This is a Plain English Papers summary of a research paper called LoCo Uses Local Preference Learning to Cut Reasoning Tokens 42–45% on Math Benchmarks. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
LoCo makes stopping a state-level decision
LoCo, or Local Compression, trains DeepSeek-R1-Distill-Qwen-7B to decide whether to continue reasoning from its current state. Its central claim is that overthinking often reflects a preference to keep generating after a correct answer is already supported, rather than an inability to solve the problem. The method therefore targets the local continue-or-conclude choice.
Instead of comparing only complete short and long solutions, LoCo freezes a reasoning prefix and compares two continuations from that identical state. The preferred continuation is regenerated, verified against the gold answer, and shorter than the original; the original continuation becomes the rejected option. This state-conditioned setup aims to shorten redundant deliberation without compressing every difficult step.
How LoCo builds local preference pairs
LoCo first screens steps in self-generated traces using average next-token confidence and its volatility across adjacent steps. It flags a state only when confidence is low and volatility is high, then asks Qwen3-30B to judge whether the next step repeats established information, makes normal progress, or gives the final answer. Only judge-confirmed overthinking states proceed....
Top comments (0)