Research map / Large language models / Instruction and fine-tuning
Towards Efficient Medical Reasoning with Minimal Fine-Tuning Data
Xinlin Zhuang, Feilong Tang, Haolin Yang, Xiwei Liu, Ming Hu, Huifa Li, Haochen Xue, Junjun He, Zongyuan Ge, Yichen Li, Ying Qian, Imran Razzak
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 |
|---|---|---|---|
| select with difficulty-influence quadrants (DIQ) instead of best baseline selector at 1% budget | Llama3.1-8B-Instruct SFT · Huatuo | Avg_A accuracy: 40.54 -> 42.78 | Improved |
| select with DIQ instead of best baseline at 10% budget | Llama3.1-8B-Instruct SFT · Huatuo | Avg_A accuracy: 41.73 -> 44.04 | Improved |
| select with DIQ instead of best baseline at 1% budget | Llama3.1-8B-Instruct SFT · FineMed | Avg_A accuracy: 39.20 -> 41.46 | Improved |
| select 10% with DIQ instead of PPL | Llama3.1-8B-Instruct SFT · FineMed | Avg_A accuracy: 38.32 -> 39.36 | Improved |
| train on 10% DIQ subset instead of full data | Llama3.1-8B-Instruct SFT · Huatuo | Avg_A accuracy: 43.44 -> 44.04 | Improved |
| train on 1% DIQ subset instead of full data | Llama3.1-8B-Instruct SFT · FineMed | Avg_A accuracy: 38.57 -> 41.46 | Improved |
| train on 1% DIQ subset instead of full data | Llama3.1-8B-Instruct SFT · Huatuo | Avg_A accuracy: 43.44 vs 42.78, -0.66 | No clear effect |
| combine difficulty and influence instead of the strongest single criterion (Reasoning-only) (ablation) | DIQ · Huatuo ablation, Llama3.1-8B-Instruct | Avg_A accuracy: +0.89 at 1%, +0.88 at 10% | Improved |
| enlarge validation set from 90 to 450 examples (ablation) | DIQ · Huatuo 1% budget | Avg_A accuracy: 42.04 -> 43.89 | Improved |
| reuse Qwen3-8B influence for Qwen3-14B/32B instead of random selection | DIQ · Qwen3-14B, Qwen3-32B | accuracy: +1.42/+1.89/+1.56 on 14B; +0.45/+2.41/+2.64 on 32B | Improved |
| transfer Llama3.1-8B influence to Qwen3 targets | DIQ · Qwen3-8B/14B/32B | accuracy: -0.01 on Qwen3-14B at 1%, -0.41 on Qwen3-32B at 1%, up to +2.11 | Mixed |
| apply 1% DIQ SFT before DPO instead of full-data SFT | Qwen3-8B SFT+DPO · FineMed preference corpus | Avg_A accuracy: 55.52 -> 56.52 | Improved |
Closest papers by meaning
- OctoMed: Data Recipes for State-of-the-Art Multimodal Medical Reasoning (CVPR 2026 · similarity 0.91)
- Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search (NeurIPS 2025 · similarity 0.89)
- Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning (CVPR 2026 · similarity 0.88)
- OpenMedReason: Scientific Reasoning Supervision for Medical Vision–Language Models (NeurIPS 2026 · similarity 0.88)
- Medical thinking with multiple images (ICLR 2026 · similarity 0.88)
- Beyond Classification Accuracy: Neural-MedBench and the Need for Deeper Reasoning Benchmarks (ICLR 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.
