LoCo: Reducing Inference Costs via Local Preference Learning
November 2, 2026
LoCo is a reasoning compression method that uses DPO to teach models when to stop deliberating. It detects overthinking states in reasoning traces and trains the model to generate shorter, verified continuations from those specific points.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference latency and token costs for reasoning models by implementing local decision points.
●
researcherThis shifts the focus of reasoning compression from trajectory-level to local decision-level optimization.