Image Captioning on COCO Caption (BLEU-1, ROUGE-L, CIDEr, SPICE, MMMU, MME, IFEval)
0.7581BLEU-1FFT
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| FFTPost-training Adaptation=Full Fine-Tuning2026.04 | 0.7581 | 0.5474 | 1.0172 | 0.2385 | 0.4244 | 735.1 | 0.207 | 0.3405 | |
| S-GRPO(wo CGI)Post-training Adaptation=Ablated Unified Post-Training without CGI2026.04 | 0.758 | 0.5473 | 1.0179 | 0.2389 | 0.5011 | 2,209.76 | 0.6067 | 0.7176 | |
| S-GRPOPost-training Adaptation=Unified Post-Training2026.04 | 0.7562 | 0.5467 | 1.0106 | 0.2385 | 0.5211 | 2,301.18 | 0.6499 | 0.7387 | |
| SFT → GRPONPost-training Adaptation=Supervised Fine-Tuning then GRPON2026.04 | 0.7501 | 0.5444 | 0.9921 | 0.2396 | 0.4522 | 901.43 | 0.3012 | 0.4123 | |
| RCPAPost-training Adaptation=Reference-based Constraint Policy Alignment2026.04 | 0.7478 | 0.5432 | 0.9814 | 0.2383 | 0.5278 | 2,289.18 | 0.647 | 0.7326 | |
| GCPOPost-training Adaptation=Gradient-based Constraint Policy Optimization2026.04 | 0.7253 | 0.5223 | 0.9649 | 0.2254 | 0.5011 | 2,003.97 | 0.5671 | 0.6732 | |
| CFFTPost-training Adaptation=Catastrophic Forgetting Free Fine-Tuning2026.04 | 0.6518 | 0.4824 | 0.7654 | 0.2034 | 0.4967 | 1,934.32 | 0.5324 | 0.6419 | |
| BASEPost-training Adaptation=Pre-trained Base Model2026.04 | 0.4457 | 0.3672 | 0.2259 | 0.1783 | 0.5122 | 2,333.36 | 0.6211 | 0.7038 | |
| GRPONPost-training Adaptation=GRPO with normalization2026.04 | 0.4437 | 0.3656 | 0.3189 | 0.179 | 0.51 | 2,315.97 | 0.6299 | 0.7238 | |
| PEFTPost-training Adaptation=Parameter-Efficient Fine-Tuning2026.04 | 0.3722 | 0.3081 | 0.1231 | 0.0862 | 0.6067 | 2,448.67 | 0.5416 | 0.6535 | |
| DAPOPost-training Adaptation=Direct Alignment Policy Optimization2026.04 | 0.2431 | 0.2798 | 0.2687 | 0.1301 | 0.5111 | 2,330.27 | 0.6577 | 0.7423 | |
| GRPOPost-training Adaptation=Group Relative Policy Optimization2026.04 | 0.2245 | 0.2767 | 0.2699 | 0.1297 | 0.5222 | 2,301.53 | 0.6506 | 0.741 |