Multimodal large language models (MLLMs) perform well after instruction tuning, but real-world deployment often demands that they keep learning from a stream of new tasks. A new paper on arXiv introduces CART, a method designed for this continual instruction-tuning setting.
The abstract notes that MLLMs achieve strong performance through instruction tuning, yet they need to continually acquire new capabilities from sequential tasks. CART is presented as a compatible routing approach, presumably to mitigate interference between tasks. The paper is titled "CART: Compatible Routing for Multimodal Continual Instruction Tuning" (arXiv:2610.22137).
Because only the abstract is available, the exact mechanism of CART is not detailed here. The significance is that it addresses a practical bottleneck for MLLMs in dynamic environments.