AIM: Anchor Identity Features, Then Match for Multimodal Large Language Model Unlearning
AIM targets identity deletion without needing retain images at deletion time.
The paper says identity facts and visual-perception questions separate in fine-tuned hidden states, giving the method room to suppress one without flattening the other. AIM first anchors a forgetting target with a universal visual prompt, then matches the vision encoder to that target under a Fisher-based constraint. The authors report competitive identity forgetting while preserving non-deleted identities, prior knowledge, and visual perception on the same images. ArXiv · AI/CL/LG's note
The paper says identity facts and visual-perception questions separate in fine-tuned hidden states, giving the method room to suppress one without flattening the other. AIM first anchors a forgetting target with a universal visual prompt, then matches the vision encoder to that target under a Fisher-based constraint. The authors report competitive identity forgetting while preserving non-deleted identities, prior knowledge, and visual perception on the same images. ArXiv · AI/CL/LG's note
score 5