---
id: 20260922-T0-12
title: "限制模型能看到什么反而学得更好？递归语言模型展现域外泛化"
title_en: "Recursive Language Models Generalize Out of Domain by Restricting Context"
url: https://ai.daily.yangsir.net/daily/20260922-T0-12
issue_date: 2026-09-22
publish_date: 2026-09-21T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2609.20831
---

# 限制模型能看到什么反而学得更好？递归语言模型展现域外泛化

研究者对比了标准思维链（CoT）与递归语言模型——后者将每个子任务放在隔离上下文中求解，限制模型能看到的范围。论文以arXiv:2609.20831发布，核心发现是：限制可见上下文反而能提升学习效果，递归语言模型在域外任务上表现出泛化能力。论文未给出具体基准名称或提升数字。该结果对长上下文模型的训练策略有参考意义：更多上下文未必等于更好推理。

## English Version

**Recursive Language Models Generalize Out of Domain by Restricting Context**

Researchers compared standard chain-of-thought (CoT) with recursive language models, which solve each subtask in an isolated context, limiting what the model can see. Posted as arXiv:2609.20831, the core finding is that restricting visible context can improve learning, with recursive models generalizing out of domain. The paper does not name specific benchmarks or report improvement figures. The result has implications for long-context training: more context does not necessarily mean better reasoning.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2609.20831)

**详情页**：https://ai.daily.yangsir.net/daily/20260922-T0-12

---

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