GOAL Targets Harmless-Query Refusals Through Gradient and Representation Changes
A new paper examines why safety-aligned language models reject harmless prompts, reporting increasing overlap between harmful and over-refusal representations in deeper layers and gradient interactions that favor refusal...