Research map / Large language models / Agents, code and math
MedAgentGym: A Scalable Agentic Training Environment for Code-Centric Reasoning in Biomedical Data Science
Ran Xu, Yuchen Zhuang, Yishan Zhong, Yue Yu, Zifeng Wang, Xiangru Tang, Hang Wu, May Dongmei Wang, Peifeng Ruan, Donghan Yang, Tao Wang, Guanghua Xiao and 4 more
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 |
|---|---|---|---|
| add SFT on successful gpt-4.1-mini trajectories (ablation) | Qwen2.5-7B-Instruct · MedAgentGym avg | average score: 16.89 -> 53.87 | Improved |
| add DPO after SFT (ablation) | Qwen2.5-7B-Instruct · MedAgentGym avg | average score: 53.87 -> 59.90 | Improved |
| use GRPO instead of DPO after SFT (ablation) | Qwen2.5-7B-Instruct · MedAgentGym avg | average score: 59.90 -> 62.17 | Improved |
| use GRPO instead of PPO after SFT (ablation) | Qwen2.5-14B-Instruct · MedAgentGym avg | average score: 69.56 -> 71.42 | Improved |
| use GRPO instead of PPO after SFT (ablation) | Qwen2.5-7B-Instruct · MedAgentGym avg | average score: 57.96 -> 62.17 | Improved |
| train with SFT or DPO on in-distribution tasks, evaluate OOD (ablation) | Qwen2.5-7B-Instruct · external OOD average | average score: 20.60 -> 20.91 (SFT), 20.84 (DPO) | No clear effect |
| train with GRPO, evaluate OOD (ablation) | Qwen2.5-7B-Instruct · external OOD average | average score: 20.60 -> 36.63 | Improved |
| train with DPO, evaluate OOD (ablation) | Qwen2.5-14B-Instruct · external OOD average | average score: 27.92 -> 27.91 | No clear effect |
| remove debugging capability from interactive coding (ablation) | Med-Copilot · MedAgentGym tasks | success rate | Improved |
| increase sampled trajectories from K=1 to K=16 (ablation) | Med-Copilot verifier · MedAgentGym | Pass@K: 17.0 -> 45.0 | Improved |
| add iterative DPO rounds after rejection-sampling SFT (ablation) | Qwen2.5-7B-Instruct · MedAgentGym | success rate: diminishing returns | Mixed |
| use coding-specialized or medical-reasoning LLMs instead of general LLMs | LLM zero-shot · MedAgentGym avg | average score: HuatuoGPT-o1-7B 16.11 vs Qwen2.5-7B-Instruct 17.43 | Hurt |
Closest papers by meaning
- BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks (NeurIPS 2026 · similarity 0.88)
- MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making (NeurIPS 2024 · similarity 0.88)
- MedMemoryBench: Benchmarking Agent Memory in Personalized Healthcare (NeurIPS 2026 · similarity 0.88)
- SciAgentGym: Benchmarking Multi-Step Scientific Tool-Use in LLM Agents (ICML 2026 · similarity 0.87)
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.87)
- MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning (ICLR 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.
