Reproduce “Think Less, Locally: Local Preference Learning for Reasoning Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang, Jie Zhou
LoCo trains a reasoning model to choose whether to continue or conclude by comparing shorter correct continuations with original continuations from the same reasoning state. It uses detector-selected overthinking points and direct preference optimization. On AIME24, DeepSeek-R1-Distill-Qwen-7B reports 54.17±4.27% accuracy with 7,364±399 tokens, versus the base model’s 50.83±6.61% and 13,493±1,367 tokens; these results measure generated-token use, not latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.