CoMMa introduces a decentralized LLM‑agent framework where specialists operate on partitioned clinical data streams. Instead of sharing raw inputs, the design enforces data decentralization to achieve stronger role specialization and applies agent‑specific fine‑tuning for performance gains. To enable interpretable coordination, CoMMa uses a contribution‑aware aggregation: deterministic embedding projections approximate each agent’s marginal utility, replacing stochastic narrative reasoning and providing explicit credit assignment and a stable decision pathway that meets clinical interpretability needs. The system is evaluated on several oncology benchmarks, including real multidisciplinary tumor board datasets from large academic hospitals in North America and East Asia as well as public datasets, demonstrating robust performance and generalization across heterogeneous clinical settings.
Review