MEXA is a learning-free framework that combines pre-trained expert models to perform scalable multimodal inference across diverse input modalities and complex tasks. To achieve effective multimodal inference across diverse domains, such as medical diagnosis and financial forecasting, MEXA dynamically selects expert models based on input modality and task-specific inference requirements. Each expert model specializes in a specific modality and task pair and generates interpretable, text-based inference outputs. MEXA aggregates and infers these outputs using a large inference model (LRM) to produce a final answer. This modular design enables flexible and transparent multimodal inference across diverse domains without additional training. Across a variety of multimodal benchmarks, including video reasoning, audio reasoning, 3D understanding, and medical QA, MEXA consistently outperforms robust multimodal-based models.
