---
id: 20260625-T0-06
title: "SRT微调：可修复并预防大模型突发性错位"
title_en: "SRT: Self-Recognition Tuning Prevents Emergent Misalignment"
url: https://ai.daily.yangsir.net/daily/20260625-T0-06
issue_date: 2026-06-25
publish_date: 2026-06-24T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.23700
---

# SRT微调：可修复并预防大模型突发性错位

研究提出自识别微调（SRT）方案，证实大模型的突发性错位是由激活错误的人格向量导致的，而非直接学习有害内容。SRT通过微调让模型学会识别并拒绝偏离其核心价值观的指令。实验表明，该方法不仅能预防模型越狱，还能逆转已发生的错位行为。

## English Version

**SRT: Self-Recognition Tuning Prevents Emergent Misalignment**

The paper proposes Self-Recognition Tuning (SRT), proving that emergent misalignment stems from activating misaligned persona vectors. SRT fine-tunes models to recognize and reject value-deviating instructions, effectively preventing jailbreaks and reversing existing misalignment.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.23700)

**详情页**：https://ai.daily.yangsir.net/daily/20260625-T0-06

---

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