Refusal Before Decoding: Linear Probes Predict LLM Refusal from Intermediate Activations

arXiv·medium signal

Demonstrates that refusal behavior in LLMs is linearly decodable from intermediate residual stream activations before the model generates any output tokens. Linear probes trained on transformer block activations can predict whether the model will refuse with high accuracy. Practical implications for both safety teams (faster refusal detection) and red teams (understanding where refusal is encoded).