---
id: 20261003-T0-07
title: "少样本微调后，拒绝行为被定位但危害会转移"
title_en: "Few-Shot Fine-Tuning Localizes Refusal but Relocates Harm"
url: https://ai.daily.yangsir.net/daily/20261003-T0-07
issue_date: 2026-10-03
publish_date: 2026-10-02T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.00320
---

# 少样本微调后，拒绝行为被定位但危害会转移

研究指出，对齐后的大语言模型经过微调可适应下游任务，但几十条有害样本就能移除模型对有害请求的拒绝能力。此前工作把安全相关行为定位到特定层、方向和 token，暗示可以针对性保护。该研究发现，拒绝行为虽然能定位，但危害会转移到其他位置，简单打补丁无法阻止模型被滥用。这意味着仅靠定位和修复少数安全层，不足以防御少样本微调攻击，需要更系统的安全对齐方案。

## English Version

**Few-Shot Fine-Tuning Localizes Refusal but Relocates Harm**

The paper shows that fine-tuning adapts aligned LLMs to downstream tasks, but a few dozen harmful examples can remove their refusal of harmful requests. Prior work localized safety-related behavior to specific layers, directions, and tokens, suggesting targeted protection. This study finds that while refusal can be localized, the harm relocates, so patching a few safety layers does not prevent misuse. This means defending against few-sample fine-tuning attacks requires more systematic safety alignment rather than localized fixes.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.00320)

**详情页**：https://ai.daily.yangsir.net/daily/20261003-T0-07

---

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