PreUnlearn: Auditing Collateral Knowledge Damage Before Large Language Model Unlearning

In three linesStudy of collateral damage in LLM machine unlearning. Authors show damage propagates beyond the forget set following semantic distance gradients, and propose PreUnlearn, a pre-unlearning prediction method to audit risks before execution.

Summary generated by Claude — human-verified