SOURCE-LINKED INTELLIGENCE
AIM: Anchor Identity Features, Then Match for Multimodal Large Language Model Unlearning
Multimodal large language models (MLLMs) can memorize identity-specific facts about people in their fine-tuning data, creating privacy risks when a person requests deletion. Existing MLLM unlearning methods often assume access to retain images or ground-truth answers during deletion, which is unrealistic in many practical scenarios. We study identity unlearning when retain images are unavailable at deletion time. Our analysis shows that identity and visual-perception questions occupy distinct regions in fine-tuned hidden states and are organized differently: identity questions cluster by perso
Read original source ↗ Open in workspace
- recordType
- paper
- region
- Global
Evidence & attribution
- arXiv · AI, language, vision and robotics · 2026-08-28T13:22:22.000Z
First collected: 2026-09-21T08:02:06.831Z. This is not the publication date.