Verbalized internal states detect hidden manipulations in medical LLMs
arXiv — cs.AI ·
What to know
A cross-type arXiv study (2611.24994) reports that translating a multimodal LLM's (Gemma-3-27B) internal activations into text via a natural-language autoencoder—termed the 'thoughts channel'—caught injected attacks significantly more often than the model's answer across all eight injected conditions (p<0.05), with near-zero false positives on clean inputs. The omission result replicated on Llama-3.3-70B, and the channel flagged unsafe clinical requests before refusal decisions, requiring no retraining.
Source and evidence
arXiv — cs.AIOriginal language · und ·Original author: Frederik Hauke, Jan Clusmann, Hanna Kreutzer, Christiane Kuhl, Tianyu Han, Felix Barajas Ordonez, Sven Nebelung, Jakob Nikolas Kather, Daniel Truhn · Source: arXiv and the named paper authors. Descriptive metadata: CC0 1.0.Source licence
Why it matters by industry
- Artificial intelligence
Relevant to AI model safety and interpretability, particularly for clinical applications where hidden manipulations could bypass output-level defenses.
Reviewed summary. The complete article remains at its original source.