LoRA Cascaded Multimodal Fusion Targets Parameter-Efficient Medical Action Recognition

arXiv·low signal

arXiv 2607.11839 combines Low-Rank Adaptation with staged multimodal fusion for nurse-training action recognition on the NurViD and Nurse Training datasets, matching dataset-specific baselines with far fewer trainable parameters. A narrow but clean demonstration of LoRA-based fusion for heterogeneous modalities.