CUSE Framework for Fine-Grained Multimodal Unlearning
October 14, 2026
CUSE implements multimodal unlearning using counterfactual guidance and on-policy self-distillation to remove sensitive or copyrighted data. It identifies specific tokens requiring forgetting by analyzing probability changes in counterfactual inputs, allowing for precise token-level control during the unlearning process.
HOW THIS AFFECTS YOU
●
builderYou can use this to more precisely remove copyrighted content from MLLMs without degrading overall model performance.
●
policyThis provides a technical mechanism for enforcing data deletion and privacy compliance in multimodal models.