CART: Compatible Routing for Multimodal Continual Tuning
October 16, 2026
CART addresses catastrophic forgetting in Multimodal Large Language Models (MLLMs) by coordinating adaptation between the projector and the language model. It uses task-specific experts for both modules combined through a weighted mixture to prevent representation drift.
HOW THIS AFFECTS YOU
●
builderThis method provides a way to continuously update multimodal models without degrading existing capabilities.
●
researcherThis architecture mitigates the asymmetric adaptation problem inherent in standard MLLM instruction tuning.