Multimodal language models remove specific data without losing other details
CUSE: Fine-Grained Multimodal Unlearning with Counterfactual Guidance and On-Policy Self-Distillation
Artificial IntelligenceComputation and Language
Summary
Some big AI models that understand pictures and text can accidentally reveal secret or unsafe information they learned before. The researchers created a way to teach these models to forget only the unwanted information while keeping everything else correct. Their method cleverly checks what parts of the model's answer need to be changed and trains it continually to remember what to keep and what to erase. This helps make AI that is safer and more respectful of privacy.
What this means in practice
- •For ai product teams: Remove specific sensitive or copyrighted information from AI models' outputs without damaging other correct outputs in commercial applications.$Commercial implications: Enables selling AI services that comply with data privacy and copyright by fine-grained content removal from multimodal models.
- •For machine learning engineers: Fine-tune multimodal language models to selectively forget undesired content while preserving desired information for safer deployment.
Authors
Chenlu Ding, Jiancan Wu, Yanchen Luo, Tianyu Zhang, Yancheng Yuan, Xiang Wang, Xiangnan He
Abstract
Multimodal large language models (MLLMs) can reproduce sensitive, copyrighted, or unsafe information from their training data. Effective unlearning must remove target information while preserving non-target content, even within the same response. Existing methods typically apply uniform forgetting objectives to complete responses and rely on fixed, off-policy supervision, resulting in coarse token control and misalignment with the model's current behavior. We propose CUSE, a Counterfactual-guided Unlearning framework via on-policy Self-distillation for multimodal large language models. It first constructs counterfactual inputs by removing target-related evidence and uses the resulting changes in token probabilities to identify which tokens require stronger forgetting. It then distills token-level forgetting and retention targets on responses sampled from the current policy, with a frozen copy of the original model serving as the teacher model augmented with the corresponding forgetting- and retention-oriented privileged contexts. This design jointly determines what to forget and what to preserve. On MLLMU-Bench, CUSE reduces forget-set ROUGE-L by 33.6% while achieving 1.2% higher retain-set ROUGE-L than the strongest baseline.