PASS (Probabilistic Agentic Supernet Sampling) is a novel multimodal agent framework applicable to medical image analysis, such as chest X-ray (CXR) inference. It is proposed to address the black-box decision-making, poor multimodal integration, and inefficient pipeline problems inherent in existing tool-based agent systems. PASS adaptively samples agent workflows on a multi-tool graph to generate interpretable, probabilistically annotated decision paths. It learns task-conditional distributions to select the most appropriate tool at each layer and provides probabilistically annotated paths for subsequent audits, improving the safety of medical AI. Furthermore, it continuously condenses key findings into an evolving personalized memory and dynamically determines whether to deepen the inference path or terminate it early for efficiency. To achieve Pareto optimization between performance and cost, we designed a three-stage training procedure that includes expert knowledge pretraining, contrastive path ranking, and cost-aware reinforcement learning. We introduce CAB-E, a comprehensive benchmark for multi-level safety-critical free-form CXR inference, and perform rigorous evaluations, demonstrating that our methods significantly outperform existing methods on various benchmarks.
