MediHive: A Decentralized Agent Collective for Medical Reasoning¶
🕒 Published (v1): 2026-03-28 05:57 UTC · Source: Arxiv · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
MediHive is a decentralized multi-agent framework for medical question answering that replaces a central coordinator with a shared append-only memory pool and peer-to-peer agent interactions. Agents autonomously self-assign roles, debate conflicting evidence, and iteratively fuse peer outputs until convergence. On MedQA and PubMedQA, it outperforms centralized multi-agent baselines using only protocol design—no fine-tuning, retrieval, or external knowledge bases.
Problem¶
Centralized LLM multi-agent systems for medical reasoning suffer from three structural defects: (1) the coordinator agent is a scalability bottleneck under high query volume; (2) it is a single point of failure; (3) when one LLM instance simulates multiple roles, role leakage and knowledge disarray degrade output quality. Decentralized designs exist for general tasks (AgentNet, Symphony) but remain largely unadapted to high-stakes healthcare QA where accuracy, ethical reasoning, and uncertainty handling are critical.
Method¶
MediHive deploys \(N\) LLM agents (Llama-3.1-70B-Instruct in experiments) communicating solely through a passive, append-only shared memory pool \(M\). The pipeline has five phases:
-
Self-Evolving Role Assignment: Each agent \(A_i\) proposes an initial role \(R_{i,0}\), posts it to \(M\), then reads all peers' proposals and refines to a final role \(R_i\) optimizing three criteria: Clarity, Differentiation, and Alignment to query \(Q\).
-
Initial Analysis: Each agent independently produces \(({\rm Reasoning}_{i,1}, {\rm Ans}_{i,1}, c_{i,1})\)—a CoT trace, an answer, and a self-assessed confidence \(c_{i,1} \in [0,1]\)—posted to \(M\).
-
Disagreement Detection and Conditional Debate: Every agent independently tallies peer votes. If agreement falls below threshold \(\tau_{\rm agree}\) (e.g., 0.8), a debate of \(T_{\rm debate}\) rounds activates. Each agent posts one of three argument types per round—Rebuttal, Defense, or Proposal—without yet updating its formal answer. The debate enriches the evidence base adversarially.
-
Iterative Shared Fusion: Starting at round \(k=2\), agents read the full history (initial analyses + debate) and update \(({\rm Reasoning}_{i,k}, {\rm Ans}_{i,k}, c_{i,k})\). For \(k>2\), only the immediately prior round's peer outputs are re-read. The loop terminates when agreement exceeds \(\tau_{\rm agree}\) for two consecutive rounds or round limit \(K\) is reached.
-
Reporter Aggregation: A post-hoc Reporter—not involved in reasoning—either confirms the supermajority answer or applies a confidence-weighted vote: $\(a^* = \arg\max_{a \in V} \sum_{i=1}^{N} c_{i,k} \cdot \mathbf{1}(a_{i,k} = a)\)$
Key Contributions¶
- Decentralized MAS architecture for medical QA: no central coordinator; all role assignment, disagreement detection, and fusion are performed autonomously by peer agents via a shared memory pool.
- Self-evolving role assignment with peer-aware refinement, ensuring diverse and query-relevant specialization without external assignment logic.
- Conditional debate phase that activates only on detected disagreement, followed by an integrative iterative fusion stage that separates adversarial evidence-gathering from formal answer convergence.
- Demonstrates that protocol design alone (without specialized training, RAG, or external tools) is sufficient to outperform centralized baselines on medical QA benchmarks.
Results¶
All baselines re-implemented with Llama-3.1-70B-Instruct for fair comparison.
- MediHive — MedQA: 84.3% accuracy / 82.5 F1; PubMedQA: 78.4% / 76.8 F1; Avg: 81.4%
- Centralized multi-agent system — 77.8% / 75.3%; Avg: 76.6%
- MedAgents — 83.7% / 76.8%; Avg: 80.3%
- Multiagent Debate — 80.4% / 78.2%; Avg: 79.3%
- Best single-agent (Few-shot CoT+SC) — 74.4% / 73.8%; Avg: 74.1%
- MediHive outperforms all single-agent and multi-agent baselines on average accuracy; achieves highest MedQA accuracy across all methods.
Limitations¶
- All experiments use a single backbone (Llama-3.1-70B-Instruct); generalization across model families or smaller models is untested.
- No external retrieval, knowledge graphs, or RAG—the system is entirely dependent on LLM parametric knowledge, which may be outdated or incomplete for niche biomedical topics.
- Self-assessed confidence scores \(c_{i,k}\) are LLM-generated heuristics, not calibrated probabilities; the weighted vote aggregation inherits this calibration weakness.
- \(\tau_{\rm agree}\), \(T_{\rm debate}\), \(K\), and \(N\) are fixed hyperparameters; the paper does not provide ablations over their sensitivity or cost-accuracy trade-offs.
- Evaluation is closed-book on static QA benchmarks; real clinical deployment (streaming patient data, time constraints, regulatory requirements) is not addressed.
- Fault-tolerance claims (agent failure resilience) are architectural arguments only—no empirical failure injection experiments are reported.
Relevance to Harnesses / Meta-Harnesses¶
MediHive is a concrete instantiation of a coordination harness: a structured runtime that orchestrates multiple LLM agents through a defined sequence of phases (role assignment → analysis → debate → fusion → aggregation) without delegating control-flow decisions to any single agent. The shared append-only memory pool functions as a harness substrate—analogous to a message bus or blackboard—that decouples agent communication from agent logic. The conditional debate gate (activate-on-disagreement) is a harness-level branching mechanism, and the Reporter's post-hoc aggregation is a harness-level synthesis step that mirrors how meta-harnesses consolidate sub-agent outputs. For researchers building harnesses, MediHive's separation of adversarial evidence enrichment (debate) from formal convergence (fusion) is a reusable protocol pattern.