CoMMa: Contribution-Aware Medical Multi-Agents From A Game-Theoretic Perspective

TL;DR

CoMMa tackles oncology decision support by replacing stochastic, narrative multi-agent coordination with data-decentralized, deterministic embeddings and a contribution-aware aggregation regulated by cooperative game theory. It architecture provides a two-stage mixture (agent-wise and global) and explicit evidence attribution through Shapley-value regularization, improving stability and interpretability. Empirical results on HCC Tumorboard and MTBBench show improved accuracy and robustness over data-centralized and role-based baselines, while enabling privacy-preserving, locally deployable inference. This approach advances practical clinical AI by delivering auditable, modality-specific credit assignment in a scalable multi-agent framework.

Abstract

Recent multi-agent frameworks have broadened the ability to tackle oncology decision support tasks that require reasoning over dynamic, heterogeneous patient data. We propose Contribution-Aware Medical Multi-Agents (CoMMa), a decentralized LLM-agent framework in which specialists operate on partitioned evidence and coordinate through a game-theoretic objective for robust decision-making. In contrast to most agent architectures relying on stochastic narrative-based reasoning, CoMMa utilizes deterministic embedding projections to approximate contribution-aware credit assignment. This yields explicit evidence attribution by estimating each agent's marginal utility, producing interpretable and mathematically grounded decision pathways with improved stability. Evaluated on diverse oncology benchmarks, including a real-world multidisciplinary tumor board dataset, CoMMa achieves higher accuracy and more stable performance than data-centralized and role-based multi-agents baselines.

Figures (8)

• Figure 1: (a) Single-agent and (b) Role-based frameworks rely on stochastic narratives, which are often costly and difficult to audit. In contrast, (c) CoMMa leverages data-decentralized agent specialization and deterministic embeddings to facilitate contribution-aware optimization and precise quantitative analysis.
• Figure 2: Architectural workflow of the CoMMa framework for multi-agent oncology decision support. (a) Deterministic embedding projection module encodes text inputs into a shared embedding space using a frozen LLM. (b) Data-decentralized agent mixture module routes each embedding to its designated partition agent and processes it via a learnable MLP. (c) Contribution-aware multi-agent module aggregate agent-specific representations using an agent-decision matrix, which is regularized by a Shapley-based contribution matrix. In the answer, we quantitatively assess each agent’s contribution. Class-wise logits and decision cutoffs are transformed into percentiles.
• Figure 3: Agentic framework response comparison on HCC Tumorboard dataset. (a) Single-agent in-context learning (GPT-4.1), (b) Role-based multi-agent framework (MDAgents-Advanced), and (c) CoMMa. For the role-based multi-agent, intermediate discussions are omitted, and only the final consensus is visualized. For CoMMa, class-wise logits and cutoff thresholds are transformed into percentiles.
• Figure 4: Visualization of agent–decision weights with and without the Contribution-Aware Multi-Agent module. With the module, weights converge toward the Shapley matrix over training, suppressing single-agent dominance. Columns are normalized to sum to 1.
• Figure 5: Additional agentic framework response comparison on HCC Tumorboard dataset - #1. (a) Single-agent in-context learning (GPT-4.1), (b) Role-based multi-agent framework (MDAgents-Advanced), and (c) CoMMa. For the role-based multi-agent, intermediate discussions are omitted, and only the final consensus is visualized. For CoMMa, class-wise logits and cutoff thresholds are transformed into percentiles.