Talk presents knowledge localization for LLM capability removal

Igor Shilov presents knowledge localization and selective gradient masking for removing capabilities from LLMs. The approach aims to effectively unlearn specific knowledge while preserving other model performance.