#artificial intelligence
May 2026
Refusal Before Decoding: Detecting and Exploiting Refusal Signals in Intermediate LLM Activations
Mechanical AutoDAN is introduced, a probe-guided variant of AutoDAN that replaces full-model fitness evaluation with partial forward passes and probe-based scoring inside a genetic prompt search loop, and it is found that refusal is linearly decodable well before the final layer, indicating that safety-relevant behavio...
M. G. Collu, Riccardo Conte, A. Giaretta et al.
· arXiv.org · 1 citation