language-backbone

This paper introduces Stochastic Meta-Unlearning (SMU), a bilevel framework that uses VLM-level feedback to learn an unlearning-ready initialization for the language backbone, achieving better forget-retain trade-offs in multimodal unlearning.