Medical Image Report Generation on OpenI
0.3396BLEU-1FFT
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| FFTPost-training Adaptation=Full Fine-Tuning2026.04 | 0.3396 | 0.2399 | 0.0903 | 0.19 | 0.4111 | 1,623.35 | 0.5323 | 0.6367 | |
| S-GRPO(wo CGI)Post-training Adaptation=Ablated Unified Post-Training without CGI2026.04 | 0.3387 | 0.2396 | 0.0901 | 0.1896 | 0.5178 | 2,301.56 | 0.6176 | 0.7001 | |
| SFT → GRPONPost-training Adaptation=Supervised Fine-Tuning then GRPON2026.04 | 0.3378 | 0.2311 | 0.0896 | 0.1839 | 0.4311 | 1,710.69 | 0.5511 | 0.6501 | |
| S-GRPOPost-training Adaptation=Unified Post-Training2026.04 | 0.3376 | 0.2388 | 0.0899 | 0.1888 | 0.5211 | 2,345.43 | 0.6298 | 0.7102 | |
| RCPAPost-training Adaptation=Reference-based Constraint Policy Alignment2026.04 | 0.3342 | 0.2325 | 0.0886 | 0.1814 | 0.5011 | 2,321.4 | 0.6285 | 0.7062 | |
| GCPOPost-training Adaptation=Gradient-based Constraint Policy Optimization2026.04 | 0.2998 | 0.2289 | 0.0829 | 0.1756 | 0.4965 | 2,201.43 | 0.5876 | 0.6882 | |
| CFFTPost-training Adaptation=Catastrophic Forgetting Free Fine-Tuning2026.04 | 0.2698 | 0.1889 | 0.0813 | 0.1698 | 0.4711 | 2,100.32 | 0.5578 | 0.6719 | |
| GRPONPost-training Adaptation=GRPO with normalization2026.04 | 0.1182 | 0.1311 | 0.0003 | 0.0999 | 0.5044 | 2,315.37 | 0.6192 | 0.711 | |
| GRPOPost-training Adaptation=Group Relative Policy Optimization2026.04 | 0.1179 | 0.1309 | 0.0003 | 0.0994 | 0.5122 | 2,356.23 | 0.6248 | 0.7062 | |
| DAPOPost-training Adaptation=Direct Alignment Policy Optimization2026.04 | 0.1165 | 0.1356 | 0.0003 | 0.0998 | 0.5111 | 2,334.65 | 0.6267 | 0.7098 | |
| BASEPost-training Adaptation=Pre-trained Base Model2026.04 | 0.1155 | 0.1299 | 0.0002 | 0.0988 | 0.5122 | 2,333.36 | 0.6211 | 0.7038 | |
| PEFTPost-training Adaptation=Parameter-Efficient Fine-Tuning2026.04 | 0.0786 | 0.0977 | — | 0.0871 | 0.6111 | 2,433.56 | 0.5508 | 0.6631 |