Found first: a primary source the press has not covered yet.
A new preprint introduces LoCo, a method for compressing large-model reasoning chains. On math benchmarks it reduces chain-of-thought length by 41.7-45.4% while maintaining or improving accuracy; on code generation, by 23.4-36.3%. The paper, titled "Think Less, Locally: Local Preference Learning for Reasoning Compression," was posted to arXiv on November 3, 2026.
What the source says
The work is by Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang, and Jie Zhou; institutional affiliations are not listed in the preprint. LoCo identifies "overthinking states" in model reasoning traces, freezes the prefix at those points, generates shorter verified continuations from that prefix, and trains preference pairs via DPO. The evaluation covers five mathematical and three code generation benchmarks; LoCo achieves the best AUC across all code benchmarks tested. Training used only self-generated math traces, yet compression transfers to code generation tasks.
Why it matters
Inference cost for reasoning models is dominated by output token count. A 41.7-45.4% reduction on math tasks with no accuracy penalty is a substantial efficiency gain for any team running these models at scale. The cross-domain transfer finding, trained on math traces alone, is a signal that overthinking states may be a general property of reasoning traces. The code results are more modest at 23.4-36.3%, and were achieved without code-specific training data.