Radiology Report Generation on RadVLM MIMIC-CXR (test)
30ROUGE-LRadVLM-Thinking+RL
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| RadVLM-Thinking+RLSize=8B, Thinking=true, RL (GRPO)=true2025.12 | 30 | 59 | 56.3 | 33.6 | 25.7 | 32.9 | 0.87 | |
| RadVLM+RLSize=8B, RL (GRPO)=true2025.12 | 29.9 | 59.2 | 57 | 33.4 | 25.8 | 32.7 | 0.86 | |
| RadVLMSize=8B2025.12 | 26 | 53.3 | 49 | 30.5 | 19 | 29.1 | 1.14 | |
| Qwen3-VL+RLSize=8B, RL (GRPO)=true2025.12 | 24.7 | 55.8 | 47.4 | 22.3 | 22.5 | 25.8 | 1.05 | |
| Qwen3-VL-Thinking+RLSize=8B, Thinking=true, RL (GRPO)=true2025.12 | 23.3 | 55 | 44.4 | 22.4 | 21 | 24.7 | 1.12 | |
| CheXagent-2Size=3B2025.12 | 22.5 | 37.4 | 54.5 | 38.7 | 20.1 | 29.9 | 1.45 | |
| DeepMedix-R1Size=7B2025.12 | 22.3 | 52.8 | 48.2 | 28.3 | 18.6 | 30 | 1.23 | |
| LLaVA-RadSize=7B2025.12 | 22.2 | 48.9 | 53.3 | 39.2 | 16.8 | 28.6 | 1.34 | |
| MedGemma-ptSize=4B2025.12 | 20.7 | 47.7 | 49.8 | 32.5 | 15.5 | 21.9 | 1.37 | |
| MAIRA-2Size=7B2025.12 | 17.7 | 46.6 | 52.1 | 35.8 | 12.9 | 21.3 | 1.42 | |
| MedGemma-itSize=27B2025.12 | 15.9 | 31.3 | 47 | 31.5 | 12 | 23.3 | 1.79 | |
| Qwen3-VL-ThinkingSize=8B, Thinking=true2025.12 | 15 | 43.1 | 34.6 | 18.5 | 10.9 | 18.5 | 1.63 | |
| Qwen3-VLSize=8B2025.12 | 14 | 42 | 35.3 | 20.4 | 10.8 | 21.9 | 1.67 |