---
id: 20261002-T0-09
title: "Hermes：让模型学会上下文推理，测试时扩展效果大幅提升"
title_en: "Hermes: Contextual Reasoning Unlocks Better Test-Time Scaling for LLMs"
url: https://ai.daily.yangsir.net/daily/20261002-T0-09
issue_date: 2026-10-02
publish_date: 2026-10-01T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.38332
---

# Hermes：让模型学会上下文推理，测试时扩展效果大幅提升

一项新研究提出Hermes方法，通过让模型学习上下文推理能力来优化测试时扩展（test-time scaling）的效果。核心思路是解决多上下文窗口场景下的决策问题：如何分配新上下文、在窗口之间传递哪些信息。该方法旨在让推理阶段额外投入的计算资源发挥更大效用。对于需要长上下文推理的任务，这一方法有望提升模型表现。

## English Version

**Hermes: Contextual Reasoning Unlocks Better Test-Time Scaling for LLMs**

A new paper introduces Hermes, a method that improves test-time scaling by teaching models contextual reasoning. The key challenge addressed is how to allocate fresh context windows and decide what information to carry between them when using additional compute during inference. The approach aims to make better use of test-time compute for tasks requiring long-context reasoning.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.38332)

**详情页**：https://ai.daily.yangsir.net/daily/20261002-T0-09

---

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