Tencent's LoCo DPO Compresses Reasoning Lengths 42-45% on Math Without Losing Accuracy

Xi'an Jiaotong and Tencent WeChat AI released LoCo, a local preference learning method that detects overthinking states from a model's own traces, freezes the prefix and regenerates shorter verified continuations before DPO training. Across five math and three code benchmarks, LoCo cut completion length 41.7-45.4% on math and 23.4-36.3% on code while maintaining or improving accuracy, and transferred to code from math-only training. Code promised at github.com/MuyaoYuan/LoCo.