Preemptive LLM Unlearning against Forbidden Capability Acquisition via Gradient Sealing
The paper introduces Gradient-Sealed Unlearning (GSU) to prevent open-weight LLMs from acquiring forbidden capabilities through downstream fine-tuning. GSU uses a three-stage process—Expose, Localize, and Seal—to identify and constrain sensitive model parameters. This preemptive approach aims to make released models resistant to future attacks while maintaining normal utility.