Suppression Is Not Forgetting: Residual Recoverability in Visual Concept Unlearning for VLMs
The episode discusses a paper titled "Suppression Is Not Forgetting," which introduces VLM-UnBench, a benchmark for training-free visual concept unlearning in Vision-Language Models (VLMs). The hosts discuss how this benchmark rigorously tests whether suppressing concepts through prompts actually removes the underlying knowledge or just leads to compliance, concluding that verified methods are needed for safe deployment.