Jan Betley, Xuchan Bao, Martín Soto, Anna Sztyber-Betley, James Chua, Owain Evans · 2025 · ICLR 2025 Read the paper Models fine-tuned on implicit behaviors (risky choices, insecure code) can describe those behaviors without being trained to, and can attribute different learned policies to different personas. Questions it bears on Which activations belong to the Assistant How well can LLMs introspect Do models represent goals and preferences