New metric audits whether AI models truly forgot sensitive information

Original: Measuring the Depth of LLM Unlearning via Activation Patching

Source: arxiv.org ↗