●builderYou can choose between fusion-based or LLM-instruction-based multimodal integration depending on your available labeled data.
●researcherThe findings suggest that the optimal multimodal architecture depends heavily on your specific training-label budget.