Research map / Large language models / Retrieval and knowledge
Small Agent Group is the Future of Digital Health
Yuqiao Meng, Luoxi Tang, Dazheng Zhang, Rafael Brens, Elvys J. Romero, Nancy Guo, Safa elkefi, Zhaohan Xi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace single 70B/72B LLM with Small Agent Group of small agents (pre-trained) | single giant LLM · MedQA, MedMCQA, NEJM-MedQA, GPQA-Bio | accuracy: MedQA 72.0 -> 86.0 (Qwen); 59.8 -> 84.6 (Llama) | Improved |
| add GRPO optimization to SAG | SAG (pre-trained) · GPQA-Bio, Qwen backbone | accuracy: 73.3 -> 92.6 | Improved |
| add CTDE optimization to SAG | SAG (pre-trained) · NEJM-MedQA, Llama backbone | accuracy: 64.9 -> 86.7 | Improved |
| add reasoning agents (AR) (ablation) | SAG · GPQA-Bio, Llama backbone | accuracy: 34.7 -> 70.1 | Improved |
| add knowledge agents with RAG (AK) (ablation) | SAG · GPQA-Bio, Llama backbone | accuracy: 45.6 -> 70.1 | Improved |
| add judgment agents (AJ) (ablation) | SAG · GPQA-Bio, Llama backbone | accuracy: 47.2 -> 70.1 | Improved |
| add safety agents (AS) (ablation) | SAG · GPQA-Bio, Llama backbone | accuracy: 59.0 -> 70.1 | Improved |
| use multi-agent debate instead of majority voting (ablation) | SAG · GPQA-Bio, Llama backbone | accuracy: 44.0 -> 70.1 | Improved |
| use SAG instead of single Qwen-72B for demographic fairness | single LLM · EquityMedQA | avg CDR: 2.4% -> 1.4% | Improved |
| use SAG under decoding variation instead of single giant LLM | single giant LLM · NEJM-MedQA consistency (Cand. setting, Llama) | std over 5 runs: ±5.47 -> ±2.12 | Improved |
| mix Llama and Qwen backbones in SAG instead of homogeneous groups (ablation) | SAG · NEJM-MedQA, 10 agents, CTDE | accuracy: 85.6 -> 91.3 (vs Qwen-only) | Improved |
| use SAG instead of single giant LLM for deployment | single 72B LLM · peak memory, latency, compute | memory / latency / TFLOPs: memory 79.5-96.8 GB vs 130 GB+; latency 58.7-63.9 s vs 27-28 s; 28.6-31.4 vs 23-25 TFLOPs | Mixed |
Closest papers by meaning
- MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks (NeurIPS 2025 · similarity 0.89)
- MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.88)
- Why Specialist Models Still Matter: A Heterogeneous Multi-Agent Paradigm for Medical Artificial Intelligence (ICML 2026 · similarity 0.87)
- PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments (NeurIPS 2026 · similarity 0.86)
- MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 2026 · similarity 0.86)
- CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework (ICLR 2026 · similarity 0.85)
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.
