ConceptGuard: Benchmarking Context-Sensitive Unlearning in Large Language Models
Existing unlearning benchmarks fail because they treat facts independently, missing how models handle concepts used for both good and bad things. ConceptGuard introduces a benchmark testing unlearning at the concept level using dual-use concepts. Results show methods like SimNPO and RMU best separate harmful and benign uses of concepts, leading to better safety while keeping useful knowledge intact.