1 article
Fed-MA’s trick is freezing 90% of the model—vision encoder and LLM—while federating only the cross-modal projector’s training.