LoCo cuts math reasoning traces 41.7-45.4% with no accuracy loss
TL;DR
- LoCo reduces completion length by 41.7-45.4% on five math benchmarks and 23.4-36.3% on three code benchmarks, per the paper.
- The method detects overthinking states in a model's own traces, freezes prefixes, regenerates shorter continuations, and trains with DPO on those pairs.
- Training data came entirely from self-generated mathematical traces, yet the authors report LoCo achieves the best AUC across all three code benchmarks.
A new preprint claims a reasoning-compression method called LoCo reduces completion length by 41.7-45.4% on mathematics benchmarks and 23.4-36.3% on code benchmarks, while "maintaining or improving accuracy," according to the paper on arXiv.
The method works at the decision level inside a trace. LoCo "detects candidate overthinking states from the model's own traces, freezes the corresponding reasoning prefixes, regenerates shorter verified continuations from the same states, and trains on the resulting state-conditioned preference pairs with DPO," the authors write. The framing is explicit: teach the model when to stop, not just to produce shorter outputs overall.
Training data came "entirely from self-generated mathematical traces," yet the authors report LoCo "achieving the best AUC across all code benchmarks" — eight benchmarks total, five math and three code.
The abstract names no base models and publishes no per-benchmark accuracy numbers. Large reasoning models, it notes, "often incurs substantial inference costs by continuing to deliberate after sufficient information has been obtained."
Originally reported by paper
Read the original article →Original headline: LoCo Cuts Frontier Reasoning Length 41-45% on Math With No Accuracy Loss