Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning

A recent study published on arXiv (2605.05909) presents a novel unlearning technique for Multimodal Large Language Models (MLLMs), which enables selective erasure of specific visual knowledge while maintaining all textual knowledge and non-target visual information. This method involves freezing the backbone of the LLM and only fine-tuning the visual component. It introduces a Contrastive Visual Forgetting (CVF) mechanism that differentiates between the targeted and preserved visual knowledge within the feature space, employing null space constraints to ensure that unlearning is confined to the retained knowledge. This innovation tackles the difficulty of managing knowledge removal and retention in MLLMs that integrate visual and textual modalities.