arXiv cs.CL· Frederik Hauke, Jan Clusmann, Hanna Kreutzer, Christiane Kuhl, Tianyu Han, Felix Barajas Ordonez, Sven Nebelung, Jakob Nikolas Kather, Daniel Truhn·· 10 小时前AI 评分58
研究:用语言化思维通道监测医疗 LLM 的隐蔽操纵
What the model knows but does not say: monitoring medical LLMs through verbalized thoughts
AI 导读
研究提出用自然语言自编码器把多模态 LLM(Gemma-3-27B)的激活转成文本,让一个看不到输入的评判模型分别读取这一“思维通道”和模型回答。在两种临床攻击目标、两种输入模态和两种投递方式共八种注入条件下,思维通道发现注入的频率均显著高于回答通道(p<0.05),两者在干净输入上的误报率都接近零;遗漏病理这一结果在 Llama-3.3-70B 上同样成立。
正文 · 原文
Abstract:Efforts to keep clinical language models safe have concentrated on what enters and leaves the model, but a manipulated model can return an answer that looks entirely ordinary. Here we show that the model's own internal state, rendered in plain language, exposes manipulations the output conceals. Using a natural-language autoencoder, we translate the activations of a multimodal LLM (Gemma-3-27B) into text and let an input-blind judge read this verbalized internal state (the thoughts channel) and the model's answer separately. Across two clinical attack objectives (omitting a pathology, covertly recommending a drug), two input modalities, and two delivery mechanisms, the thoughts channel caught the injection significantly more often than the answer in all eight injected conditions (p<0.05), while both channels held near-zero false-positive rates on clean inputs; the omission result carried over to Llama-3.3-70B. The thoughts channel also registered unsafe clinical requests as harmful while the model was still reading them, before it decided whether to refuse. None of this requires retraining: an off-the-shelf verbalizer and a general-purpose judge are enough.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.24994 [cs.CL] |
| (or arXiv:2611.24994v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.24994 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Frederik Hauke [view email]
[v1]
Wed, 18 Nov 2026 14:15:51 UTC (2,467 KB)
来源:arXiv cs.CL · arxiv.org