# Think Less, Locally: Local Preference Learning for Reasoning Compression

**Authors:** Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang, Jie Zhou  
**Lab:** Tencent · **Published:** 3 Nov 2026 · **Upvotes:** 73 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.04814) · [PDF](https://arxiv.org/pdf/2611.04814) · [Code](https://github.com/MuyaoYuan/LoCo) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.04814.html)

## Curator's notes

**TL;DR:** LoCo learns when to stop reasoning locally, cutting chain length while keeping accuracy.

- **What's new:** It trains models on state‑conditioned preference pairs to decide continuation versus conclusion, unlike prior methods that only shorten whole trajectories.
- **Key result:** LoCo reduces completion length by 41.7%–45.4% on math and 23.4%–36.3% on code while maintaining or improving accuracy.
- **Why it matters:** Builders can deploy cheaper, faster reasoning models without sacrificing performance.

## Abstract

Large reasoning models achieve strong performance on challenging tasks, but their long chain-of-thought often incurs substantial inference costs by continuing to deliberate after sufficient information has been obtained. Existing reasoning compression methods primarily operate on reasoning trajectories, paying less attention to the local decision of whether further reasoning is necessary given the current state. We propose LoCo, a local preference learning method for reasoning compression. 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. This formulation teaches the model when to continue reasoning and when to conclude, rather than simply encouraging shorter trajectories overall. We construct training data entirely from self-generated mathematical traces and evaluate LoCo on five mathematical and three code generation benchmarks. LoCo reduces completion length by 41.7\%--45.4\% on mathematics and 23.4\%--36.3\% on code, while maintaining or improving accuracy and achieving the best AUC across all code benchmarks. LoCo consistently improves the accuracy-efficiency trade-off on mathematics and transfers effectively to code generation despite math-only training, demonstrating the cross-domain generality of local preference learning for reasoning compression. The code and model will be publicly released at https://github.com/MuyaoYuan/LoCo.

## What you could build

1. Implement a math‑solver API that stops early using LoCo’s decision model
2. Create a code‑completion tool that trims generation time with LoCo
3. Build a benchmark suite comparing inference cost versus accuracy with and without LoCo

## Quiz yourself

What does LoCo train the model to predict instead of merely encouraging shorter trajectories?

**Answer:** It trains the model to predict whether to continue reasoning or to conclude given the current reasoning state.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.