---
id: 20261111-T0-03
title: "不改权重也能无限上下文？新方法让现有模型记住所有对话"
title_en: "Unbounded Context Added to Stock LLMs Without Retraining Weights"
url: https://ai.daily.yangsir.net/daily/20261111-T0-03
issue_date: 2026-11-11
publish_date: 2026-11-10T05:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2611.10141
---

# 不改权重也能无限上下文？新方法让现有模型记住所有对话

这篇论文提出一种方法，不改模型权重，只通过调整模型的几何结构（旋转注意力等），就让两个现成语言模型具备了“永不遗忘”的能力：读过的每份文档、说过的每一轮对话都能保留，并据此作答。研究发布在arXiv（编号2611.10141），属于新增论文。实际意义在于，现有模型不用重新训练、不用换架构，就有望获得无限上下文能力，长文档问答和多轮对话场景可以直接受益。

## English Version

**Unbounded Context Added to Stock LLMs Without Retraining Weights**

A new arXiv paper (2611.10141) shows that two stock language models can retain every document they have read and every turn they have spoken, then answer from any of it, without modifying their weights. The trick is manipulating the models' geometry, including rotary attention. The upshot: existing models could gain unbounded context without retraining or architecture changes, a direct win for long-document QA and multi-turn dialogue.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2611.10141)

**详情页**：https://ai.daily.yangsir.net/daily/20261111-T0-03

---

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