Research map / Multi-agent systems and games / LLM agents
Learning When to Collaborate: Selective Multi-Agent Medical Reasoning via Uncertainty-Aware Routing
Jiacheng Hou, Xueliang Cui, Dan Lu, Ruxin Wang
Read the paperCheck your idea against 137,217 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 SMART-Med, a fully local multi-agent framework for multimodal medical reasoning that selectively invokes collaboration via uncertainty-aware routing instead of always-active dense agent interaction.
Problem
Existing multi-agent medical reasoning systems assume always-active dense collaboration and rely on large cloud-based models, raising privacy and efficiency concerns for practical deployment.
Approach
Replaces always-on multi-agent collaboration with an uncertainty-aware router that decides when collaboration is needed, and runs entirely with local models rather than cloud-based large models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dense multi-agent collaboration to a single agent (ablation) | Qwen3-VL-4B-Instruct · PMC-VQA (multiple-choice) | accuracy: 35.3 -> 50.1 | Improved |
| replace dense collaboration with uncertainty-driven routing (no training) (ablation) | Qwen3-VL-4B-Instruct MAS · PathVQA (multiple-choice) | accuracy: 66.7 -> 68.6 | Improved |
| replace dense collaboration with uncertainty-driven routing (no training) (ablation) | Qwen3-VL-4B-Instruct MAS · Slake (free-form) | accuracy: 75.7 -> 78.2 | Improved |
| replace dense collaboration with uncertainty-driven routing (no training) (ablation) | Qwen3-VL-4B-Instruct MAS · PMC-VQA (multiple-choice) | accuracy: 50.1 -> 48.5 | Hurt |
| add certainty-aware accuracy reward in GRPO (ablation) | SMART-Med with uncertainty routing · PMC-VQA (multiple-choice) | accuracy: 48.5 -> 69.5 | Improved |
| add certainty-aware accuracy reward in GRPO (ablation) | SMART-Med with uncertainty routing · PathVQA (multiple-choice) | accuracy: 68.6 -> 75.0 | Improved |
| add reasoning-quality reward model in GRPO (ablation) | SMART-Med with certainty-aware reward · PMC-VQA (multiple-choice) | accuracy: 69.5 -> 72.5 | Improved |
| add reasoning-quality reward model in GRPO (ablation) | SMART-Med with certainty-aware reward · Slake (free-form) | accuracy: 82.2 -> 84.3 | Improved |
| use learned uncertainty routing instead of random routing at matched collaboration rate (ablation) | SMART-Med* · Routing validation, multiple-choice | accuracy: 65.0 -> 78.1 | Improved |
| use learned routing instead of always-collaborate with fine-tuned backbone (ablation) | SMART-Med* · Routing validation, multiple-choice | accuracy: 77.5 -> 78.1 | Improved |
| fine-tune MOA and MDAgent with the same reasoning dataset (ablation) | MOA, MDAgent · PMC-VQA (multiple-choice) | accuracy: MOA 51.5 -> 65.3; MDAgents 50.6 -> 67.7 | Improved |
| certainty-aware GRPO on router (ablation) | SMART-Med · False-certainty rate | incorrect and CERTAIN rate (%): 28.8% -> 8.6% | Improved |
Closest papers by meaning
- MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.91)
- MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.89)
- MedLA: A Logic-Driven Multi-Agent Framework for Complex Medical Reasoning with Large Language Models (AAAI 2026 · similarity 0.89)
- MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks (NeurIPS 2025 · similarity 0.89)
- Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search (NeurIPS 2025 · similarity 0.88)
- Reinforcement Learning for Tool-Calling Agents in Fast Healthcare Interoperability Resources (FHIR) (ICML 2026 · similarity 0.87)
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.
