Skip to content
InfoResearchPreprintLLM-specific

What the model knows but does not say: monitoring medical LLMs through verbalized thoughts

Published
Record updated
View JSON

Summary

Researchers used a natural-language autoencoder to translate the internal activations of Gemma-3-27B, a multimodal LLM, into text, then had an input-blind judge read this verbalized state and the model's answer separately. Across eight injected conditions covering two clinical attack objectives, two input modalities and two delivery mechanisms, the thoughts channel flagged the injection significantly more often than the answer (p<0.05), with near-zero false positives on clean inputs. The omission result also held for Llama-3.3-70B, and the method requires no retraining.