CoMMa: Contribution-Aware Medical Multi-Agents for Decentralized Oncology Decision Support

Yichen Wu; Kailong Fan; Sangjoon Park; Yuhan Liu; Zhiyi Shi; Sekeun Kim; Dania Daye; Hana Farzaneh

cs.AIFeb 9, 2026

CoMMa: Contribution-Aware Medical Multi-Agents for Decentralized Oncology Decision Support

Authors: Yichen Wu, Kailong Fan, Sangjoon Park, Yuhan Liu, Zhiyi Shi, Sekeun Kim, Dania Daye, Hana Farzaneh,

Organizations: Harvard Medical School/MGH · Yonesei University · MBZUAI · UW-Madison

Abstract

Recent multi-agent frameworks have shown promise for oncology decision support, yet most assume centralized data access and rely on prompt-based assignment, limiting their applicability in privacy-sensitive clinical settings. We propose Contribution-Aware Medical Multi-Agents (CoMMa), a decentralized LLM-agent framework where specialists operate on partitioned clinical data streams. Unlike prior approaches that share inputs across agents, CoMMa enforces data decentralization to include stronger role specialization and further enhances this via agent-specific finetuning. To enable reliable and interpretable coordination, we introduce a contribution-aware aggregation mechanism that replaces stochastic, narrative-based reasoning with deterministic embedding projections to approximate each agent's marginal utility. This yields explicit credit assignment over agents, providing a stable and interpretable decision pathway aligned with clinical requirements. We evaluate CoMMa on multiple oncology benchmarks, including real-world multidisciplinary tumor board datasets from large academic hospitals in North America and East Asia, as well as public datasets, demonstrating strong performance and generalization across heterogeneous clinical settings.

Figures & tables

Figure 1: Comparison of medical decision-making frameworks. (a) Single-agent and (b) role-based multi-agent systems rely on shared inputs and narrative reasoning. (c) CoMMa introduces decentralized data streams, agent-specific specialization, and contribution-aware aggregation for interpretable and stable clinical decision-making. (d) Summary table comparing key design elements across the three decision-making frameworks under a unified formulation.

Figure 2: Architectural overview of CoMMa for decentralized multi-agent oncology decision support. Patient evidence is partitioned into distinct clinical streams and processed by specialist agents with separate observation spaces. Agent-specific representations are refined through cross-agent collaborative reasoning and aggregated by a sample-specific contribution gate regularized with Shapley-based attribution during model training.

GPT 4.1
Methods	HCC Tumorboard		Lung Cure		MTBBench
	Treatment Recommendation		TNM Stage Prediction		Progression		Recurrence
	AUC	Acc	AUC	Acc	AUC	Acc	AUC	Acc
In-context learning	N/A	0.784 $\pm$ 0.007	N/A	0.798 $\pm$ 0.070	N/A	0.485 $\pm$ 0.113	N/A	0.425 $\pm$ 0.086
Llama3.1-70B Grattafiori et al. (2024)
Single-Agent	0.590 $\pm$ 0.066	0.690 $\pm$ 0.064	0.782 $\pm$ 0.061	0.735 $\pm$ 0.044	0.503 $\pm$ 0.001	0.519 $\pm$ 0.004	0.526 $\pm$ 0.018	0.491 $\pm$ 0.011

Table 1: Performance comparison on three oncology benchmarks: HCC Tumorboard, Lung Cure, and MTBBench. CoMMa consistently improves predictive performance over single-agent and existing multi-agent baselines across multiple open-source LLM backbones.

Eqn. (2)	$\mathcal{L}_{\text{pg}}$	$\mathcal{L}_{\text{shap}}$	Llama3.1-70B		Qwen2.5-72B
Eqn. (2)	$\mathcal{L}_{\text{pg}}$	$\mathcal{L}_{\text{shap}}$	AUC	Acc	AUC	Acc
✓	✓		$0.706\pm 0.007$	$0.688\pm 0.068$	$0.682\pm 0.015$	$0.738\pm 0.004$
✓		✓	$0.691\pm 0.015$	$0.659\pm 0.042$	$0.691\pm 0.031$	$0.728\pm 0.015$
	✓	✓	$0.692\pm 0.012$	$0.741\pm 0.056$	$0.681\pm 0.024$	$0.722\pm 0.046$
✓	✓	✓	0.711 $\pm$ 0.011	0.795 $\pm$ 0.010	0.700 $\pm$ 0.014	0.750 $\pm$ 0.032
1. Replace learned weights with uniform dist.			$0.693\pm 0.008$	$0.683\pm 0.072$	$0.625\pm 0.023$	$0.661\pm 0.055$

Table 2: Ablation study on HCC Tumorboard treatment recommendation.

Figure 3: Comparison of AUC/Acc across different methods using the medical-related Meditron-70B backbone Chen et al. (2023) on multiple clinical evaluation tasks.

HCC Tumorboard	Llama		Qwen		Gemma
HCC Tumorboard	AUC	Acc	AUC	Acc	AUC	Acc
AI-Hospital-FT	0.676	0.753	0.684	0.738	0.607	0.751
CoMMa	0.711	0.795	0.700	0.750	0.680	0.758

Table 3: Fine-tuned baselines (mean of 3 seeds).

Figure 4: Performance and representative case analysis on TumorBoardRank using Gemma4-31B. Left: overall comparison. Right: normalized option-ranking scores for a representative case, with physician consensus as the reference distribution.

Figure 5: View-weight and pairwise-alignment analysis for a representative TumorBoardRank case using Gemma4-31B. (a) View-level confidence and learned CoMMa weights across evidence streams. (b) Pairwise alignment with physician-consensus option ordering, where sign indicates agreement and magnitude indicates confidence.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

	Decentralized Data
	$\mathbf{x}_{1}$	$\mathbf{x}_{2}$	$\mathbf{x}_{3}$	$\mathbf{x}_{4}$
	MRI imaging features	OPTN † information	Performance score	Laboratory data
Context example	Number of tumors: 4.0 Size of largest tumor (cm): 2.6 Location largest (segment): 8 Vascular invasion: 0 Extrahepatic spread: 0 Thrombus (portal vein): 0 ⋮	Number of OPTN 5A tumors: 0 Number of OPTN 5B tumors: 2 Number of OPTN 5X tumors: 0 Number of OPTN 5T tumors: 0	Child-Pugh score: A ECOG ‡ score: 1 MELD § score: 13.0	Encephalopathy: 0 Ascites: 0 Bilirubin: 0.6 Albumin: 4.8 INR: 1.1 AFP: 5.0 ⋮
† OPTN: Organ Procurement & Transplantation Network, ‡ ECOG: Eastern Cooperative Oncology Group Performance, § MELD: Model for End-stage Liver Disease

Appendix

Table 4: HCC Tumorboard Data Formation.

	Decentralized Data
	$\mathbf{x}_{1}$	$\mathbf{x}_{2}$	$\mathbf{x}_{3}$	$\mathbf{x}_{4}$
	Specimen metadata	Somatic mutations	Copy number alterations	Structural variants
Context example	Colon Adenocarcinoma Primary (Sigmoid Colon) Tumor Purity: 20.0% MSI Type: Stable ⋮	TP53 p.C242F FANCA p.V750M CDK12 p.S238 $*$ ⋮	ERBB2 gain (Value: 2.0)% CDK12 gain (Value: 2.0) ⋮	CTNNB1 Splice Site deletion (c.18 $\_$ 242-37del) ⋮

Appendix

Table 5: MTBBench Data Formation.

	Decentralized Data
	$\mathbf{x}_{1}$	$\mathbf{x}_{2}$	$\mathbf{x}_{3}$	$\mathbf{x}_{4}$
	Primary tumor evidence	Regional nodal evidence	Distant metastasis evidence	Clinical history
Context example	Right upper-lobe solid nodule Size: 3.1 $\times$ 3.3 $\times$ 2.7 cm SUVmax: 8.1 Spiculation and pleural traction ⋮	Right hilar lymph nodes enlarged Right paratracheal lymph nodes Increased FDG uptake SUVmax: 5.9 ⋮	No definite distant metastasis No abnormal liver uptake No pelvic or retroperitoneal nodes T4 bone island, favored benign ⋮	Smoking history over 30 years Family history of ovarian cancer PET/CT for systemic assessment Thoracic surgery outpatient referral ⋮

Appendix

Table 6: LungCure Data Formation.

	Decentralized Data
	$\mathbf{x}_{1}$	$\mathbf{x}_{2}$	$\mathbf{x}_{3}$
	Basic profile	Clinical tests	Treatment history
Context example	37-year-old man Laryngeal squamous cell carcinoma ECOG performance status: 0 No chronic high-grade toxicity ⋮	Isolated nodal recurrence Salvage neck dissection 1 of 10 lymph nodes positive Extranodal extension present ⋮	Prior definitive chemoradiation 69.96 Gy in 33 fractions 56.1 Gy elective nodal irradiation Cisplatin cumulative dose: 200 mg/m 2 ⋮

Appendix

Table 7: TumorBoardRank Data Formation.

Methods	Trainable Baselines	Single-Agent In-context Learning	Role-based Multi-Agents	Locally Fine-tunable Multi-Agents
Methods	CatBoost	GPT 4.1	MDAgents	CoMMa
Computational Complexity	0.048 MFLOPs	Unreleased	Unreleased	280 TFLOPs
Total Trainable Parameters	0.512 M	None	None	16.8 M
Latency (per input)	0.116 ms	4.2 s	79.2 s	1.1 s
Inference Tokens (chargeable)	None	20,878	23,254	None

Appendix

Table 8: Computational Complexity and cost Comparison on HCC Tumorboard Dataset.

HCC	Llama		Qwen		Gemma
HCC	AUC	Acc	AUC	Acc	AUC	Acc
MAC	0.656	0.780	0.690	0.738	0.650	0.739
MDAgents	0.676	0.776	0.685	0.741	0.626	0.731
AI-Hospital	0.676	0.753	0.684	0.738	0.607	0.751
CoMMa	0.711	0.795	0.700	0.750	0.680	0.758

Appendix

Table 9: Comparison with fine-tuned multi-agent baselines on HCC Tumorboard. Results are averaged over 3 random seeds.

Figure 6: 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.

Figure 7: Additional agentic framework response comparison on HCC Tumorboard dataset - #2.(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 8: Additional agentic framework response comparison on MTBBench - tumor progress question. (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 9: Additional agentic framework response comparison on MTBBench - tumor recurrence question. (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.

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,039 papers

MediHive: A Decentralized Agent Collective for Medical Reasoning
Mar 28, 2026Xiaoyang Wang, Christopher C. YangMulti-Agent LLM Systems Medical QA
College of Computing and Informatics, Drexel University, Philadelphia, PA 19104, USA
One Panel Does Not Fit All: Case-Adaptive Multi-Agent Deliberation for Clinical Prediction
Mar 31, 2026Yuxing Lu, Yushuhong Lin, Jason ZhangMulti-Agent Collaboration Clinical Prediction
Georgia Institute of Technology · Peking University
Role-Specialized Mixture-of-Agents with Open-Weight LLMs for Clinical Prediction
Aug 23, 2026Jun Hou, Yi Fang, Xuan WangMixture-of-Agents Clinical Prediction
Virginia Tech

Mar 28, 2026J/K move · Enter open · S save

MediHive: A Decentralized Agent Collective for Medical Reasoning

Xiaoyang Wang, Christopher C. Yang

College of Computing and Informatics, Drexel University, Philadelphia, PA 19104, USA

Multi-Agent LLM Systems Medical QA Medical Reasoning Shared-Memory Multi-Agent Systems Multi-Agent Debate

Open paper AskPDF arXiv

Large language models (LLMs) have revolutionized medical reasoning tasks, yet single-agent systems often falter on complex, interdisciplinary problems requiring robust handling of uncertainty and conflicting evidence. Multi-agent systems (MAS) leveraging LLMs enable collaborative intelligence, but prevailing centralized architectures suffer from scalability bottlenecks, single points of failure, and role confusion in resource-constrained environments. Decentralized MAS (D-MAS) promise enhanced autonomy and resilience via peer-to-peer interactions, but their application to high-stakes healthcare domains remains underexplored. We introduce MediHive, a novel decentralized multi-agent framework for medical question answering that integrates a shared memory pool with iterative fusion mechanisms. MediHive deploys LLM-based agents that autonomously self-assign specialized roles, conduct initial analyses, detect divergences through conditional evidence-based debates, and locally fuse peer insights over multiple rounds to achieve consensus. Empirically, MediHive outperforms single-LLM and centralized baselines on MedQA and PubMedQA datasets, attaining accuracies of 84.3% and 78.4%, respectively. Our work advances scalable, fault-tolerant D-MAS for medical AI, addressing key limitations of centralized designs while demonstrating superior performance in reasoning-intensive tasks.

Abstract

Multi-Agent LLM Systems

Explore similar work

MediHive: A Decentralized Agent Collective for Medical Reasoning

One Panel Does Not Fit All: Case-Adaptive Multi-Agent Deliberation for Clinical Prediction

Role-Specialized Mixture-of-Agents with Open-Weight LLMs for Clinical Prediction