Dongxi 东锡 NLP· @dongxi_nlp · X·· 2026-08-07AI 评分22
AI 导读
模型可以在不暴露其影响的情况下被引导。 悄无声息的轻推可以逃过推理监控器的检测。 论文： 《在隐式影响场景下，思维链监控可能不可靠》
正文 · 原文
A model can be steered without revealing the influence.
Quiet nudges can escape reasoning monitors.
Paper:
Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings
New paper, led by Agatha Duzan: your CoT monitorability numbers are too optimistic. Models are bad at hiding on demand: instructions to do a bad thing leaks into the CoT. How much does this inflate the numbers? Our new benchmark measures it 🧵在 X 查看被引用的帖子
来源：Dongxi 东锡 NLP · x.com