Reasoning models can be trained to stop when they know the answer

A new method teaches large language models when to conclude reasoning rather than continue, cutting output length by up to 45% on math and 36% on code without losing accuracy.

Chinese Tech

Muyao Yuan · Liyan Xu · Weizhan Zhang · Muyan Jiao · Jiangnan Li · Jinchao Zhang · +1 more

Xi’an Jiaotong University · Tencent Inc.

Research Digest·4 November 2026·3 min read

Illustration: Reasoning models can be trained to stop when they know the answer

In one line

LoCo trains reasoning models to stop when they have enough information, reducing generated text while maintaining or improving accuracy on math and code tasks.

Before you read this

Two ideas this paper builds on

preference optimizationA post-training method that adjusts model behavior using comparisons or self-generated preferences to improve alignment.57 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.291 papers →

Read it as

Some AI systems write out long chains of thought, meaning step-by-step reasoning, even after they have enough information to answer. This wastes time and computing power, and extra steps do not always improve answers.

LoCo finds points in a model’s own worked solutions where it may already have enough information. From each point, it compares a shorter, checked answer with a longer continuation, then trains the model to choose when to stop and when to keep reasoning.

The method shortened generated text by roughly two-fifths on math tasks and about a third on code tasks, while maintaining or improving accuracy. Code improvements came even though training used only math examples, suggesting the learned stopping behavior can transfer across tasks.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

reasoning compression local preference learning preference optimization chain of thought cost-quality trade-off verification