CRAM: Centroid-Routing Adaptive MoE Prevents Catastrophic Forgetting in Multimodal Continual Instruction Tuning

arXiv·medium signal

Addresses catastrophic forgetting when MLLMs learn new vision-language tasks sequentially by introducing centroid-routing and adaptive Mixture-of-Experts. New tasks get dedicated expert capacity while routing preserves access to prior knowledge. For teams that need to continually fine-tune multimodal models on new domains without retraining from scratch, CRAM offers a practical architecture that scales with task count.