Research map / Large language models / Agents, code and math
MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MDAgents, an adaptive multi-agent framework that classifies a medical query's complexity (low/moderate/high) and assigns a solo, multi-disciplinary team, or integrated care team structure accordingly, evaluated on ten medical QA and multimodal benchmarks.
Problem
Determining the best way to use LLMs for complex medical decision-making, since static solo or group multi-agent setups may be suboptimal or costly across tasks of varying complexity.
Approach
An LLM moderator classifies each medical query's complexity and the framework recruits a single primary care clinician, a multi-disciplinary team, or an integrated care team accordingly, with a moderator review and optional MedRAG knowledge to refine group outputs.
Key results
- MDAgents achieved best accuracy on 7 of 10 benchmarks, e.g. 88.7% on MedQA and 77.9% on DDXPlus versus the best baselines.
- Adding moderator review and MedRAG together raised average accuracy from 71.8% to 80.3% (+11.8%) in the ablation.
- On the full MedQA 5-option set with GPT-4o mini, MDAgents reached 83.6% versus 80.2% for Reconcile.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt collaboration structure to query complexity rather than use the strongest single-agent baseline | MDAgents · MedQA | accuracy (%): 83.9 -> 88.7 | Helped |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · PubMedQA | accuracy (%): 79.7 -> 75.0 | Hurt |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · Path-VQA | accuracy (%): 62.1 -> 65.3 | Helped |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · PMC-VQA | accuracy (%): 53.4 -> 56.4 | Helped |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · MedVidQA | accuracy (%): 57.8 -> 56.2 | Hurt |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · DDXPlus | accuracy (%): 72.7 -> 77.9 | Helped |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · SymCat | accuracy (%): 91.9 -> 93.1 | Helped |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · JAMA | accuracy (%): 71.0 -> 70.9 | No clear effect |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · MedBullets | accuracy (%): 77.0 -> 80.8 | Helped |
| adapt collaboration structure to query complexity rather than use the strongest baseline | MDAgents · MIMIC-CXR | accuracy (%): 53.5 -> 55.9 | Helped |
| add MedRAG (ablation) | MDAgents · All datasets | average accuracy (%): 71.8 -> 75.2 | Helped |
| add moderator's review (ablation) | MDAgents · All datasets | average accuracy (%): 71.8 -> 77.6 | Helped |
Surprising findings
- Solo methods outperformed group settings on MedQA, and group settings were not uniformly better across benchmarks.
- Increasing the number of agents in the group setting did not improve performance; the adaptive method peaked with N=3 agents.
Limitations
- Evaluation is restricted to multiple-choice question answering rather than interactive, patient-centered diagnosis.
- Multi-agent adaptive approach is far more expensive in API cost than solo prompting.
- Experiments use only 50 test samples per dataset with some benchmarks showing large variance across seeds.
Experimental setup
Datasets: MedQA PubMedQA DDXPlus SymCat JAMA MedBullets Path-VQA PMC-VQA MIMIC-CXR MedVidQA MedRAG
Benchmarks: MedQA PubMedQA DDXPlus SymCat JAMA MedBullets Path-VQA PMC-VQA MIMIC-CXR MedVidQA
Metrics: accuracy API calls consensus entropy
Baselines: Zero-shot Few-shot CoT CoT-SC Ensemble Refinement Medprompt Majority Voting Weighted Voting Borda Count MedAgents Meta-Prompting ReConcile
Builds on: chain-of-thought self-consistency Medprompt ReConcile MedAgents
Closest papers by meaning
- MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks (NeurIPS 2025 · similarity 0.92)
- Why Specialist Models Still Matter: A Heterogeneous Multi-Agent Paradigm for Medical Artificial Intelligence (ICML 2026 · similarity 0.90)
- MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.90)
- MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow (ICLR 2026 · similarity 0.89)
- MedChain: Bridging the Gap Between LLM Agents and Clinical Practice with Interactive Sequence (NeurIPS 2025 · similarity 0.89)
- MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.