InfoResearchPreprintLLM-specific
What the model knows but does not say: monitoring medical LLMs through verbalized thoughts
- Published
- Record updated
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.
Related items
- InfoAnthropic shares fall after blockbuster opening as wider tech sell-off continuesSame vendor · BBC Technology
- MediumBazuka Android Malware Uses Fake Screen Overlays to Hide Credential Theft and Device ControlSame vendor · The Hacker News
- InfoAnthropic raises $93.5 billion total as underwriters scoop up ‘greenshoe’ overallotmentSame vendor · CNBC Technology
- LowIn Other News: Norwegian Energy Firm Cyberattacks, Supercomputer Hack, New Nightmare Eclipse ExploitSame vendor · SecurityWeek
- LowSignal is blocking Gemini and Siri AI from reading your screenSame vendor · The Verge (AI)