MedPrune: Topology-Efficient Multimodal Multi-Agent Communication Evolution for Medical VQA Tasks

What it claims

MedPrune keeps the underlying multimodal LLMs frozen and learns only the communication graph between specialist agents. It learns which agents to drop and which links to cut. Across MIMIC-CXR, IU-Xray and OmniMedVQA, in both full-set and few-shot training settings, the authors report:

• an average performance gain of 3.65% over training-free, training-intensive and medical multi-agent baselines;
• a 37.86% improvement in token efficiency;
• an average performance drop of only 3.1% under adversarial attack.

The introduction does not say which metric the 3.65% refers to, which baseline it is measured against, or how token efficiency is defined.

How it works

The diagnostic dialogue is a graph. Each node is a role-prompted specialist agent assigned to a department, such as respiratory medicine, ophthalmology or cardiology. There are two edge types:

• Spatial edges connect agents within the same round.
• Temporal edges connect agents across rounds.

Each edge type is further split into intra-department and inter-department versions. Every agent updates its memory from the question plus the weighted messages from its neighbours. After T rounds, a summarizer agent produces the answer.

1. Soft adjacency. The binary adjacency matrices are relaxed with Gumbel-Softmax into soft matrices in [0,1]. There is a separate matrix for each round and each edge type. Message weights are a softmax over each agent's neighbourhood.
2. Node pruning (one-shot). The soft matrices are trained with REINFORCE-style policy gradient. K graphs are sampled and task performance is the reward. Policy gradient is used because the MLLM "may be accessed via non-differentiable APIs." Then, in each round, the nodes with the lowest summed in- and out-degree are removed along with all their edges, at pruning rate α. The point is to remove token cost from round 1 onward, not to refine a topology gradually.
3. Edge pruning. Edge weights are re-initialized on the reduced graph. The intra- and inter-department matrices are then optimized separately with the same policy gradient. The objective is task reward minus the nuclear norm of each matrix, which pushes toward low-rank, sparse graphs. A Frobenius-norm bound δ limits how far the soft matrices can drift from the original binary graph. Low-utility edges are then pruned progressively.

Compared with adaptive-topology methods, the authors' claimed difference is efficiency. Those methods iterate topology refinement and still spend many tokens. MedPrune cuts agents up front and needs fewer rounds (T2 ≪ T1). Among the prior work it cites, the closest is M3Prune, which prunes only edges in a hierarchical communication graph. MedPrune adds node pruning and department-aware heterogeneous edges.

How to use it

No code is available yet. The paper says source code and data are in the supplementary material and "will be released upon acceptance." The method needs task-labelled VQA data to compute rewards, because the graph is trained in full-set and few-shot settings. It also needs K sampled multi-agent rollouts per gradient step. The available text gives no figure for this training cost, and does not name the backbone MLLM.

Limits

• The fetched text stops partway through the method section (at "Progressive Edge Pruning"). Results tables, ablations, the backbone model, the choices of α, K, δ and T, and the attack setup were all not visible. Every number above comes from the introduction's summary.
• The 37.86% token-efficiency figure is reported without a stated definition. It is also unclear whether it includes the token cost of training the graph.
• The evaluation covers three VQA benchmarks. Two of them, MIMIC-CXR and IU-Xray, are chest X-ray datasets. Nothing in the visible text tests the method on clinical deployment.
• This is v1 of a preprint released 2026-10-05.