FASD Method for Honest Ignorance in MLLM Unlearning
October 14, 2026
Forget-Aware Self-Distillation (FASD) decouples knowledge suppression from refusal learning to prevent hallucination and superficial refusals. The method aims for honest ignorance, ensuring a model genuinely loses target knowledge and can acknowledge its absence rather than just declining queries.
HOW THIS AFFECTS YOU
●
researcherYou can evaluate unlearning effectiveness beyond surface-level refusals by testing for underlying knowledge retention.
●
policyThis helps ensure models actually forget sensitive data rather than just learning to hide it behind refusals.