---
id: 20260702-T0-03
title: "LearnStop：推理模型何时该停手？一种无隐藏状态的早退方案"
title_en: "LearnStop: A Cost-Aware Study on When Reasoning Models Should Stop"
url: https://ai.daily.yangsir.net/daily/20260702-T0-03
issue_date: 2026-07-02
publish_date: 2026-07-01T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2606.30852
---

# LearnStop：推理模型何时该停手？一种无隐藏状态的早退方案

arXiv 论文《When Does Learning to Stop Help?》提出了一种名为 LearnStop 的推理模型早退方案。研究发现，不同的推理实例需要不同的计算量，简单的置信度阈值往往不是最优解。LearnStop 是一种无隐藏状态的方法，旨在让模型学会在最佳时机停止计算，以平衡准确性与推理成本。

## English Version

**LearnStop: A Cost-Aware Study on When Reasoning Models Should Stop**

The paper 'When Does Learning to Stop Help?' proposes LearnStop, a hidden-state-free method for early exits in reasoning models. It suggests that learned stopping rules can better balance accuracy and cost compared to simple confidence or convergence thresholds.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2606.30852)

**详情页**：https://ai.daily.yangsir.net/daily/20260702-T0-03

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*