MedPrune prunes both specialist agents and their connections
MedPrune is a medical multimodal multi-agent collaboration framework that dynamically prunes nodes and edges from the communication topology. The paper formulates the diagnostic process as a heterogeneous communication graph: nodes are specialist agents drawn from various departments, and edges capture intra- and inter-departmental interactions.
Two sparsification mechanisms drive this. Heterogeneous Node Sparsification removes specialist agents irrelevant to the current multimodal question using reinforcement learning-driven topological optimization. Heterogeneous Edge Sparsification then selectively retains only the most diagnostically salient intra- and inter-departmental connections by jointly optimizing task performance and topological complexity. The stated goal of the two-stage design is to reach high diagnostic performance with drastically reduced token costs.
Redundant topologies waste tokens before any answer is produced
The paper frames the problem around existing clinical workflow-inspired multi-agent frameworks, which it says suffer from interaction patterns and excessive computational overhead caused by redundant communication topologies. In the authors' account, existing implementations often rely on static, predefined interaction workflows, and this rigidity fails to capture the dynamic, case-adaptive nature of real-world clinical consultations while token consumption escalates as more agents and discussion rounds are involved.
The paper's own comparison figure distinguishes three behaviors. Fixed-Template Evolving relies on static, predefined communication templates and degrades diagnostic performance through noisy redundant discussions. Adaptive Evolving can also reach a smaller communication subgraph in the final stage, but its iterative topology refinement still incurs substantial token overhead. MedPrune instead prunes irrelevant agents and redundant edges simultaneously, which the authors describe as token-efficient multi-department diagnostic reasoning with fewer rounds.
Node pruning removes departments the question does not need
The paper argues that existing multi-agent medical systems typically communicate with all department agents, compelling models to allocate computational resources to clinically irrelevant specialties. Heterogeneous Node Sparsification addresses this by learning a task-adaptive inter-departmental communication graph.
Mechanically, the heterogeneous graph is parameterized with learnable soft adjacency matrices, and these are optimized via reinforcement learning using task performance as the reward signal. Nodes with minimal weighted degree, along with their incident edges, are then pruned. The stated effect is to dynamically assemble a compact team of relevant specialists and eliminate token overhead from the initial round. The paper's illustrative example shows a chest X-ray pleural effusion question where non-radiology and non-respiratory agents reply that the question is not related to them.
Edge pruning keeps only diagnostically salient connections
Even after irrelevant agents are pruned, the paper states that redundant communication edges may persist, wasting tokens and propagating hallucinations across departments. Heterogeneous Edge Sparsification therefore re-initializes edge weights on the node-sparsified subgraph and jointly optimizes them under a composite objective that maximizes task performance while penalizing topological complexity.
This process prunes low-utility edges and retains only the most salient intra- and inter-departmental connections, so that each remaining edge contributes high diagnostic utility. The paper describes the combination of node and edge pruning as a two-stage sparsification that fundamentally addresses efficiency bottlenecks in multi-agent medical reasoning.
Accuracy rises 3.65% while token efficiency rises 37.86%
MedPrune is evaluated on three medical VQA benchmarks: MIMIC-CXR, IU-Xray, and OmniMedVQA, under full-set and few-shot training settings. The paper reports that the method consistently outperforms strong baselines across training-free, training-intensive, and medical multi-agent systems.
Under both training settings, the paper reports an average performance improvement of 3.65% and a 37.86% boost in token efficiency. These two numbers measure different things and should not be conflated: one is an average accuracy-side gain across the evaluated benchmarks, the other is a relative efficiency gain in token use. In the few-shot table with 20 and 40 training samples on a Qwen3-VL (8B) backbone, MedPrune reaches an average of 75.98 and 75.97 respectively, against 71.74 for MedAgent-Pro and 71.27 for M3Prune at 40 samples. The paper states the improvements are statistically significant at p < 0.05.
Robustness holds under adversarial attacks and prompt variants
The paper reports that MedPrune reduces average performance by only 3.1% under adversarial attack settings, which the authors present as evidence of the system's robustness. This is a drop relative to the method's own non-attacked performance, not a comparison against baseline robustness.
Prompt robustness is tested separately. Across prompt variants, the paper reports only minor performance differences, with the largest deviation below 1.1%, which the authors say confirms that MedPrune's learned topology is robust to prompt changes without retraining. The variants include a prompt-simplified version with only core instructions and a prompt-reversed version that fully swaps the role prompts of chief physicians and regular doctors.
Transfer holds, but scaling beyond 15 agents is untested
Cross-dataset transfer is evaluated under full-data training with a shared Qwen3-VL (8B) backbone, in two scenarios: training on MIMIC-CXR and testing on unseen IU-Xray, and training on PMC-OA and testing on unseen MIMIC-CXR. The paper reports that all methods experience moderate performance drops due to dataset shift, yet MedPrune consistently outperforms the baselines, which the authors take as evidence that the node-edge pruning strategy generalizes to unseen datasets.
The limitations qualify how far this extends. The sparsification framework involves key hyperparameters such as the node dropout rate and the noise tolerance, and the paper states these values are not exhaustively optimized across all possible departmental compositions and imaging modalities; optimal pruning aggressiveness may vary with question complexity and the number of collaborating specialists. Department selection currently uses a frozen MLLM relevance-scoring heuristic rather than being integrated into the reinforcement learning optimization. The framework scales effectively with up to 15 agents across 3 departments, but the paper does not evaluate much larger cohorts of subspecialists, leaving the tractability of policy gradient optimization under expanded combinatorial action spaces untested.
Sources used in this answer
MedPrune: Topology-Efficient Multimodal Multi-Agent Communication Evolution for Medical VQA Tasks
MedPrune prunes both specialist nodes and communication edges in a heterogeneous multi-agent graph, reporting an average 3.65% performance improvement and a 37.86% token efficiency boost across three medical VQA benchmarks under full-set and few-shot training.