NTH

Think Less, Locally: Local Preference Learning for Reasoning Compression

AuthorsMuyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang, Jie Zhou

AffiliationsMOEKLINNS, School of Computer Science and Technology, Xi’an Jiaotong University · WeChat AI, Tencent Inc

November 8, 2026 2 min read
Watch on YouTube
The one-line take

LoCo teaches reasoning models when to stop thinking, cutting response length while preserving or improving accuracy.

Key results

45.4%
AIME24 completion-length reduction

Reduction relative to the base model.

54.17%
AIME24 accuracy

LoCo accuracy; the base model scores 50.83%.

36.3%
MBPP token reduction

Reduction relative to the base model after math-only training.

23.4%
CodeContests token reduction

Reduction relative to the base model after math-only training.

What the paper found

LoCo teaches a reasoning model when to stop, rather than simply rewarding shorter answers. It detects likely overthinking states in the model’s own traces using low confidence and high confidence volatility, then asks a Qwen3-30B judge to confirm redundancy. From each confirmed state, it freezes the prefix, generates eight continuations, selects the shortest one verified against the answer, and uses the original continuation as the rejected option in DPO training. Trained on math-only traces from DeepSeek-R1-Distill-Qwen-7B, LoCo reduces completion length while preserving or improving accuracy across five math and three code benchmarks, and achieves the best AUC on all three code benchmarks. On AIME24, it cuts completion length by 45.4% and reaches 54.17% accuracy, compared with 50.83% for the base model. Despite training only on math, it reduces tokens by 36.3% on MBPP and 23.4% on CodeContests. Ablations, including one using GPT-5.4-generated continuations, support the value of learning local preferences from the model’s own reasoning states.

Original 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.

Read the original paper

More in AI Reasoning

Browse all 51 papers →
01Reasoning

Scaling Lean Theorem Proving for Real-World Code Verification

Debargha Ganguly, Divya Raghunathan, Dejan Jovanović, Vikash Singh, Soonho Kong, Aaron Tomb, Aman Goel, Ferhat Erata, Sam Bayless

EasyProver tackles Lean proofs at scale and introduces harder verification challenges grounded in real Rust code.

Read analysis