---
id: 20260522-T0-15
title: "基于代理的长上下文推理方法：仅需部分输入"
title_en: "Proxy Context Boosts Long-Document Reasoning Efficiency"
url: https://ai.daily.yangsir.net/daily/20260522-T0-15
issue_date: 2026-05-22
publish_date: 2026-05-21T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2605.20201
---

# 基于代理的长上下文推理方法：仅需部分输入

arXiv论文提出代理上下文方法，解决百万token输入下的推理效率问题。当前大模型在长文本复杂推理任务中表现不佳，研究发现仅需使用输入子集（代理上下文）即可完成任务。该方法通过训练模型识别关键信息片段，在保持95%推理准确率的同时减少70%计算量。在合同条款分析等场景中已验证有效性。

## English Version

**Proxy Context Boosts Long-Document Reasoning Efficiency**

arXiv paper proposes proxy context method to improve reasoning efficiency with long inputs. Current models perform poorly on complex long-context tasks despite supporting 1M+ tokens. Research shows only input subsets (proxy contexts) are needed for reasoning. By training to identify key information segments, it maintains 95% accuracy while reducing computation by 70%. Validated in contract clause analysis scenarios.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2605.20201)

**详情页**：https://ai.daily.yangsir.net/daily/20260522-T0-15

---

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