LoCo learns when to stop reasoning
A new paper proposes LoCo, a DPO-based method that detects candidate overthinking states in a reasoning model’s traces, freezes the corresponding reasoning prefixes, regenerates shorter verified continuations and trains on state-conditioned preference pairs. The authors report completion-length reductions of 41.7%–45.4% on five mathematical benchmarks and 23.4%–36.3% on three code-generation benchmarks, while maintaining or improving accuracy.
The paper addresses the inference costs caused by continued deliberation after sufficient information has been obtained. LoCo makes state-dependent decisions about whether to continue reasoning or conclude, rather than simply shortening every reasoning trajectory.
Computer Science > Artificial Intelligence
arXiv:2611.04814 (cs)
[Submitted on 3 Nov 2026]
Title:Think Less, Locally: Local Preference Learning for Reasoning Compression
Authors:Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang, Jie Zhou
View a PDF of the paper titled Think Less, Locally: Local Preference Learning for Reasoning Compression, by Muyao Yuan and 6 other authors