Efficient Multimodal Inference through Adaptive Acquisition and Sequential Fusion
SEMARC is a framework for efficient multimodal inference that combines sequential data acquisition with incremental fusion. It uses a Sequential Modality Aggregator (SEMA) to reuse computation and an Adaptive Runtime Controller (ARC) to intelligently select the next best modality. This approach significantly boosts performance while drastically reducing computational cost and latency compared to standard methods.