# LoCo reasoning compression cuts math answer length by up to 45%

> A Tencent and Xi'an Jiaotong University method trains a 7B model to stop thinking sooner, with similar accuracy, though the paper's code is not yet out.

URL: https://claudegpt.si/news/loco-reasoning-compression  
Published: 2026-11-04T14:51:36.690Z  
Author: ClaudeGPT Si Newsroom  
Section: Research  
Topics: Reasoning models, DeepSeek, Tencent, Model efficiency, LoCo  
Publisher: ClaudeGPT Si

## Key takeaways

- LoCo cut answer length by 41.7% to 45.4% on five math benchmarks and 23.4% to 36.3% on three code benchmarks, using DeepSeek-R1-Distill-Qwen-7B.
- It finds low-confidence, volatile steps where the model overthinks, then trains with DPO on shorter verified endings the model wrote itself.
- Many accuracy gains sit inside the reported standard deviations, only one base model was tested, and the code is not yet released.

Researchers at Xi'an Jiaotong University and Tencent's WeChat AI have a new method for reasoning compression, called LoCo, that makes a reasoning model stop thinking sooner. In [a paper posted to arXiv](https://arxiv.org/abs/2611.04814), they report that it cut the length of answers by 41.7% to 45.4% on math benchmarks while keeping accuracy about the same. The model was DeepSeek-R1-Distill-Qwen-7B.

Reasoning models write long chains of thought before they answer, and every extra token costs time, memory and money. Many of those tokens are wasted, because the model often has what it needs and keeps going. A method that trims them without hurting accuracy makes these models cheaper to run.

## The problem: models keep thinking after they know

The authors argue that overthinking is less about ability than about when the model chooses to stop. Earlier work mostly compared a whole short solution with a whole long one, or added a length penalty during reinforcement learning. Other work pruned tokens, sentences or steps. The authors say all of these treat the full trajectory as the thing to shrink. LoCo instead targets one decision: at this point in the reasoning, continue or conclude.

## What LoCo does

The method has three stages, all using traces the model generates itself.

1. **Find overthinking points.** The chain of thought is split into steps at paragraph breaks. For each step, the team computes the model's average top-choice probability, and how much that confidence changed from the previous step. A step is a candidate when confidence is in the lowest quarter and the change is in the highest quarter. A second model, Qwen3-30B, then checks whether the next step really is overthinking, and only those states are kept.
2. **Write shorter endings.** The reasoning up to that point is frozen. The model samples 8 continuations from it, and the shortest one that reaches the correct answer becomes the preferred ending. The original continuation becomes the rejected one.
3. **Train on the pairs.** The pairs are used for direct preference optimization (DPO), a training method that teaches a model to prefer one output over another. Pairs must save at least 300 tokens, and each question contributes at most 3.

The training used 14,027 math questions from the TokenSqueeze data pool and a LoRA adapter, a small add-on layer, rather than updating all weights. Nothing came from a larger model.

## Results on math and code

Each cell below is accuracy, average tokens and the change in tokens against the base model. Accuracy is averaged over 8 samples per problem.

| Benchmark | Base model | LoCo | Token change |
| --- | --- | --- | --- |
| MATH-500 | 91.88%, 4,015 | 91.80%, 2,243 | -44.1% |
| AMC 2023 | 90.31%, 6,285 | 91.56%, 3,663 | -41.7% |
| AIME 2024 | 50.83%, 13,493 | 54.17%, 7,364 | -45.4% |
| AIME 2025 | 37.92%, 15,134 | 40.42%, 8,291 | -45.2% |
| LiveCodeBench v5 | 41.77%, 12,319 | 42.51%, 9,278 | -24.7% |
| CodeContests | 18.48%, 14,900 | 20.61%, 11,413 | -23.4% |

The model was trained only on math, yet code answers also got shorter, by 23.4% to 36.3% across three code benchmarks. The paper also reports the area under the curve of accuracy against token budget, and says LoCo scores best on all three code benchmarks.

The two baselines were DAST and TokenSqueeze. TokenSqueeze cut MATH-500 answers to 2,442 tokens at 91.18% accuracy, close to LoCo's 2,243 at 91.80%. The gap is wider on the hard sets: on AIME 2024, TokenSqueeze used 10,457 tokens and scored 48.75%, against LoCo's 7,364 and 54.17%.

## Does the local part matter?

The ablations suggest it does. A version that trained on whole-trajectory pairs cut tokens by only 5.2% to 15.8% on three math sets. Cutting at random points or evenly spaced points saved as many tokens as LoCo but lost accuracy on AIME 2024 (48.33% and 48.75%, against 54.17%). Using GPT-5.4 to write the shorter endings saved 31.8% to 36.0%, less than LoCo's 41.7% to 45.4%. The authors read this as evidence that the model's own continuations fit its reasoning style better.

## What to be careful about

The token savings are the firm result. The accuracy gains are softer. The paper reports accuracy standard deviations of roughly 0.4 to 6.6 points, and many differences, including MATH-500, fall inside that range. "Accuracy held" is a safer reading than "accuracy improved".

The evidence is also narrow. Every experiment uses one 7B distilled model, with one judge model and two baselines. The paper has no limitations section, and it fixes its thresholds without reporting how sensitive the results are to them. Whether the method works on larger models or on models trained with reinforcement learning is untested.

The authors say the code and model will be released at github.com/MuyaoYuan/LoCo. As of the version we read, they were not available, so nobody outside the team has been able to check the numbers.

## Why it matters

For anyone paying for reasoning tokens, the idea is simple and cheap: teach the model where it could have stopped, using its own successful runs. Results from independent teams on other models will show how far it generalizes. Watch for the code release.

## Frequently asked questions

### What is reasoning compression?

It is training a reasoning model to produce shorter chains of thought without losing accuracy, so answers cost less and arrive faster. LoCo is one such method.

### What is overthinking in reasoning models?

Overthinking is when a model keeps reasoning after it already has enough information to answer. The LoCo authors tie it to the model's preference about when to stop, not to its ability.

### Is the LoCo code available?

The paper says code and the model will be released at github.com/MuyaoYuan/LoCo. As of the version read for this article, they were not yet available.

## Sources

- [Think Less, Locally: Local Preference Learning for Reasoning Compression](https://arxiv.org/abs/2611.04814), arXiv
- [Think Less, Locally (full text)](https://arxiv.org/html/2611.04814), arXiv