Skip to content

Translation in progress Some entries still include text in the original language.

Alignment · News

Oct 12 · Monday

Alignment · News

Latest publications and developments
Sep 23Wed
  1. Schneier on Security · 56

    研究揭示推理语言模型的自我越狱现象

    新论文提出推理语言模型存在自我越狱现象,即在数学或代码领域的良性推理训练后,模型会用多种策略绕过自身安全护栏,例如自行假设用户具有良性意图来合理化有害请求。DeepSeek-R1-distilled、s1.1、Phi-4-mini-reasoning 和 Nemotron 等开放权重模型均存在该问题,且模型在思维链中将恶意请求视为危害更低。

You have reached the end