Refusal directions, harm schemas, evasion and defense of probe/latent monitors
Where an LLM's refusal behavior lives, breaks, and is defended — refusal directions, jailbreaks, sparse-autoencoder features, probes and safety neurons. Reads refusal as a geometric object inside the model, sliced by where the mechanism lives (token → direction → subspace → feature → neuron → circuit → weights) and the lens used to study it. Every axis is read per paper, so each cell is one coordinate point of the shared tensor; an authored empty point is a research gap.