---
id: 20261120-T0-04
title: "GUARD：用遗憾引导稳住AI代理的递归自我改进"
title_en: "GUARD: Stabilizing Recursive Self-Improvement with Regret Guidance"
url: https://ai.daily.yangsir.net/daily/20261120-T0-04
issue_date: 2026-11-20
publish_date: 2026-11-19T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.23476
---

# GUARD：用遗憾引导稳住AI代理的递归自我改进

论文提出GUARD方法，解决自主AI代理递归自我改进中的稳定性问题。这类代理由冻结的基础模型加可编辑的提示词、技能、工具规则和记忆组成，能在不更新权重的情况下从经验中自我修改。但缺乏外部评分器时，自我编辑容易失控。GUARD用遗憾引导来约束这一过程。对做Agent自我进化方向的研究者来说，这是针对失控风险的一个具体方案。

## English Version

**GUARD: Stabilizing Recursive Self-Improvement with Regret Guidance**

A new paper introduces GUARD to address stability in recursive self-improvement for autonomous AI agents. These agents pair a frozen base model with an editable harness of prompts, skills, tool rules, and memory, self-modifying from experience without weight updates. Without external graders, self-editing can go off the rails. GUARD uses regret guidance to constrain the process—a concrete proposal for researchers working on agent self-evolution and its failure modes.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.23476)

**详情页**：https://ai.daily.yangsir.net/daily/20261120-T0-04

---

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