Research map / Multi-agent systems and games / LLM agents

MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

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 MMedAgent-RL, which RL-trains a triage doctor and an attending physician (Qwen2.5-VL-7B) to coordinate proprietary specialist LVLMs, using curriculum RL with dynamic entropy regulation to teach the attending physician when to trust specialist opinions.

Problem

Static GP-specialist-GP multi-agent workflows for medical multimodal VQA lack flexibility, and specialists' outputs are noisy, so the aggregating physician can copy wrong judgments; single medical LVLMs also generalize poorly across specialties.

Approach

Train the triage and attending GP agents with GRPO, and build a three-stage curriculum (easy/medium/hard) based on specialist accuracy, with a per-stage entropy bonus coefficient (near zero for easy, increasing for hard) to balance exploitation of specialist advice against correcting it.

Key results

• MMedAgent-RL (7B) averages 73.3% on in-domain VQA-RAD, SLAKE, PathVQA, the best among reported models, above GPT-4o at 68.6%.
• On OOD benchmarks MMedAgent-RL reaches 72.6% average (OmniMedVQA 73.3, MMMU-Med 71.9), versus AFlow 56.6%.
• Majority-voting test-time scaling raises the average to 76.1% in-domain and 76.6% OOD.

Surprising findings

• Using specialist accuracy as curriculum difficulty, the easy-only stage can hurt OOD performance relative to the full curriculum; the paper reports the full three-stage schedule is needed for gains.
• Plain majority voting over specialists can suppress a minority correct opinion, which motivated the entropy-aware design.

Limitations

• Specialist outputs come from proprietary models (GPT-4o, o3), which limits reproducibility and adds API cost.
• Difficulty stratification depends on ground-truth labels of specialist correctness, which are only available during training.
• Theoretical analysis relies on strong assumptions (PL condition, smoothness, local convexity) that are not verified for LLM policies.

Experimental setup

Datasets: VQA-RAD SLAKE PathVQA

Benchmarks: VQA-RAD SLAKE PathVQA OmniMedVQA MMMU (Health & Medicine)

Metrics: accuracy

Baselines: GPT-4o LLaVA-v1.6-34B HuatuoGPT-Vision-7B MedVLThinker-7B Qwen2.5-VL-7B MedAgents MDAgents AFlow

Builds on: GRPO Qwen2.5-VL curriculum learning MDAgents MedAgents

Closest papers by meaning

• MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.90)
• MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning (ICLR 2026 · similarity 0.89)
• Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning (ICML 2026 · similarity 0.89)
• CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework (ICLR 2026 · similarity 0.88)
• MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks (NeurIPS 2025 · similarity 0.88)
• Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search (NeurIPS 2025 · similarity 0.88)