ProtoAda: Prototype-Guided Adapter Expansion for Multimodal Continual Learning Without Forgetting

arXiv·medium signal

Proposes prototype-guided adaptive adapter expansion with geometric consolidation for multimodal continual instruction tuning. Dynamically grows adapter capacity based on task similarity measured via prototypes, then consolidates using geometric methods to prevent interference. Complementary to CRAM — both target the same problem (MLLM continual learning) but from different angles: adapters vs. MoE routing.