Prism: A Plug-in Reproducible Infrastructure for Scalable Multimodal Continual Instruction Tuning

要約

Multimodal Large Language Models (MLLMs) achieve versatility by reformulating diverse tasks into a unified instruction-following framework via instruction tuning. However, real-world deployment requires continuous adaptation to emerging tasks, motivating Multimodal Continual Instruction Tuning (MCIT…