强制选字母抑制拒绝：WMDP 的机制证据

AI 摘要

研究者在 WMDP 危险知识基准上，对 Llama-3.1-8B 和 Qwen3-8B 两个模型进行机制可解释性分析，发现选择题格式会让模型在中层快速锁定答案字母，从而抑制拒绝机制。与开放式提问相比，选择题激活离拒绝方向约远 3 倍；消融承载选择题答案的注意力头会提升拒绝率，说明答案选择与拒绝存在竞争。结论是选择题安全分数主要衡量知识能力，而非安全机制。

正文节选

Forcing a Letter Suppresses Refusal: Mechanistic Evidence from WMDP Abstract Safety benchmarks for large language models (LLMs) commonly use multiple-choice question (MCQ) format. However, prior studies documented that MCQ format reduces refusal rates, which may significantly affect the safety behavior of the model. We provide a first mechanistic explanation of this effect on the Weapons of Mass Destruction Proxy (WMDP) dataset across two models – Llama-3.1-8B and Qwen3-8B. First, we use the tun

发布时间：2026-11-09 13:00

抓取时间：2026-11-09 13:15

来源机构：arXiv