GRADE: Probing Knowledge Gaps in LLMs through Gradient Subspace Dynamics