ViKeR: Vision-guided key token regularization for multimodal large language model unlearning

The researchers propose a vision-guided key token regularization method, ViKeR, for unlearning in multimodal large language models. The method uses irrelevant visual inputs to predict the ideal token-level distribution after unlearning in order to regularize the unlearning process, defines key tokens by information entropy, and amplifies their updates through token-level gradient reweighting. Experiments on the MLLMU and CLEAR benchmarks show that the method can effectively unlearn target information while mitigating forgetting and maintaining response coherence.