MedPrune: Topology-Efficient Multimodal Multi-Agent Communication Evolution for Medical VQA Tasks
MedPrune trims both specialist agents and their communication links to make medical VQA collaboration cheaper and stronger.
The paper models diagnosis as a heterogeneous graph of department-style agents and their intra- and inter-department exchanges. It uses reinforcement learning to drop task-irrelevant agents, then keeps only the most diagnostically useful connections. The authors say experiments on medical VQA, including full-set and few-shot settings, beat multi-agent baselines while improving token efficiency and adversarial robustness. ArXiv · AI/CL/LG's note
The paper models diagnosis as a heterogeneous graph of department-style agents and their intra- and inter-department exchanges. It uses reinforcement learning to drop task-irrelevant agents, then keeps only the most diagnostically useful connections. The authors say experiments on medical VQA, including full-set and few-shot settings, beat multi-agent baselines while improving token efficiency and adversarial robustness. ArXiv · AI/CL/LG's note
score 4