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

Researchers introduced Prism, a lightweight plug-in framework for multimodal continual instruction tuning (MCIT).

Prism separates new algorithm development from the base MLLM codebase, reducing the need to modify core code.

The design supports scalable training workflows and easier experimentation across methods.

The goal is to lower engineering overhead, improve fairness in comparisons, and make results more reproducible.