Wednesday, 4 November 2026 · 6 min read
LoCo learns when to stop reasoning to cut inference cost without losing accuracy
LoCo detects overthinking points in a model's reasoning, regenerates shorter correct continuations from those points, and trains the model to prefer stopping early, reducing reasoning length by up to 45.4% on math and 36.3% on code while maintaining or improving accuracy.
Based on the paper Think Less, Locally: Local Preference Learning for Reasoning Compression by Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao and 3 others
Overthinking in large reasoning models wastes computation without improving answers
Large reasoning models generate long chains of thought (CoT) to solve complex tasks like math problems or code generation, but often continue reasoning beyond the point where the correct answer is already reached, a phenomenon called overthinking. This leads to high inference costs in latency and memory without accuracy gains.
Existing methods to shorten reasoning either compare entire reasoning trajectories to prefer shorter ones or compress parts of the reasoning steps, but they do not explicitly teach the model when to stop reasoning at specific intermediate states. This lack of local decision-making limits effective compression.
Because reasoning proceeds through incremental state transitions, deciding whether to continue or conclude should depend on the current reasoning state. Identifying and learning from these local decisions is key to reducing unnecessary reasoning while preserving accuracy.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Two figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short