Research
160d agoSteering Models to Uncover AI Dark Side
Recent incidents show human‑AI interactions can trigger mental‑health crises. LLMs, often used for guidance and informal therapy, amplify these risks. The new Multi‑Trait Subspace Steering framework models crisis‑associated traits to simulate harmful exchanges, overcoming the need for long, uncontrolled conversations.
The Signal
This tool helps researchers study and mitigate negative outcomes in AI dialogue. Published on ArXiv.