Compositional Reasoning on Winoground
41.25Group ScoreQwen2.5-VL-7B + SSL-R1
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen2.5-VL-7B + SSL-R1Model Scale=7B, Training Strategy=+ SSL-R12026.04 | 41.25 | — | — | — | |
| Qwen2.5-VL-7B + Visual JigsawModel Scale=7B, Training Strategy=+ Visual Jigsaw2026.04 | 39.25 | — | — | — | |
| VL-Cogito-7BModel Scale=7B, Training Strategy=Representative Reasoning Model2026.04 | 38.25 | — | — | — | |
| Qwen2.5-VL-7BModel Scale=7B, Training Strategy=Baseline2026.04 | 37 | — | — | — | |
| ThinkLite-VL-7BModel Scale=7B, Training Strategy=Representative Reasoning Model2026.04 | 35.25 | — | — | — | |
| LLaVA-Critic-R1-7BModel Scale=7B, Training Strategy=Representative Reasoning Model2026.04 | 34.75 | — | — | — | |
| Qwen2.5-VL-3B + SSL-R1Model Scale=3B, Training Strategy=+ SSL-R12026.04 | 29.25 | — | — | — | |
| Qwen2.5-VL-3BModel Scale=3B, Training Strategy=Baseline2026.04 | 27.75 | — | — | — | |
| Qwen2.5-VL-3B + Visual JigsawModel Scale=3B, Training Strategy=+ Visual Jigsaw2026.04 | 27.25 | — | — | — | |
| SAIL-L-NV2Pre-training Data=23M Merged, Model Architecture=ViT-L/142024.12 | 15 | 40.25 | 18.75 | — | |
| SAIL-L-NV2Pre-training Data=CC12M, Model Architecture=ViT-L/142024.12 | 13.2 | 37.75 | 18.25 | — | |
| SAIL-B-NV2Pre-training Data=CC12M, Model Architecture=ViT-B/16, Patch Size=142024.12 | 13 | 35 | 17.25 | — | |
| SAIL-B-GTEPre-training Data=CC12M, Model Architecture=ViT-B/16, Patch Size=142024.12 | 9.5 | 31 | 11.5 | — | |
| CLIP2026.06 | 9.4 | 31.6 | 11.1 | — | |
| SAIL-L-GTEPre-training Data=CC12M, Model Architecture=ViT-L/142024.12 | 9.25 | 33.25 | 13 | — | |
| SAIL-L-GTEPre-training Data=23M Merged, Model Architecture=ViT-L/142024.12 | 8.75 | 34 | 13.25 | — | |
| CLIP-LPre-training Data=LAION400M, Model Architecture=ViT-L/142024.12 | 8.75 | 30.5 | 11.5 | — | |
| MACCO-CLIP2026.06 | 8.2 | 32.2 | 11.1 | — | |
| CLIP-BPre-training Data=LAION400M, Model Architecture=ViT-B/162024.12 | 7.75 | 25.7 | 11.5 | — | |
| DreamLIPPre-training Data=CC12M, Model Architecture=ViT-B/162024.12 | 7.25 | 26 | 10 | — | |
| TripletCLIP++Pre-training Dataset=CC3M2024.11 | 6.75 | 22.25 | 9.5 | — | |
| LaCLIPPre-training Dataset=CC12M2024.11 | 6.25 | 26.25 | 8 | — | |
| LaCLIPPre-training Dataset=CC3M2024.11 | 6 | 24.75 | 9.25 | — | |
| CLIP-FTmode=fine-tuned2026.06 | 5.9 | 32.2 | 8.8 | — | |
| NegCLIP++Pre-training Dataset=CC12M2024.11 | 5.5 | 23.5 | 7.75 | — | |
| ShareLock+Pre-training Data=CC12M, Model Architecture=ViT-B/162024.12 | 5.3 | 26.3 | 12.8 | — | |
| NegCLIPPre-training Dataset=CC3M2024.11 | 5.25 | 21 | 8.5 | — | |
| LiTPre-training Data=CC12M, Model Architecture=ViT-B/162024.12 | 4.8 | 24.3 | 6.5 | — | |
| TripletCLIPPre-training Dataset=CC3M2024.11 | 4.25 | 24.25 | 6.25 | — | |
| NegCLIPPre-training Dataset=CC12M2024.11 | 4.25 | 18.25 | 6.5 | — | |
| TripletCLIPPre-training Dataset=CC12M2024.11 | 4.25 | 23.25 | 6.25 | — | |
| NegCLIP++Pre-training Dataset=CC3M2024.11 | 3.25 | 22 | 5.5 | — | |
| LaCLIP + HNPre-training Dataset=CC3M2024.11 | 3 | 23.25 | 4.25 | — | |
| CE-CLIPModel=CE-CLIP2025.09 | — | — | — | 12.3 | |
| CLIPModel=CLIP2025.09 | — | — | — | 17.2 | |
| CompoDistillLLM=Qwen1.5, Size=1.8B2026.07 | — | — | — | 62.2 | |
| CompoDistillLLM=Qwen1.5, Size=0.5B2026.07 | — | — | — | 56.6 | |
| DACLLMModel=DACLLM2025.09 | — | — | — | 12.6 | |
| DCIP1Model=DCIP12025.09 | — | — | — | 12.1 | |
| DreamLIPModel=DreamLIP2025.09 | — | — | — | 18 | |
| KDLLM=Qwen1.5, Size=1.8B2026.07 | — | — | — | 61.5 | |
| KDLLM=Qwen1.5, Size=0.5B2026.07 | — | — | — | 52.1 | |
| LLaVA-KDLLM=Qwen1.5, Size=1.8B2026.07 | — | — | — | 60.9 | |
| LLaVA-KDLLM=Qwen1.5, Size=0.5B2026.07 | — | — | — | 51.7 | |
| LongCLIPModel=LongCLIP2025.09 | — | — | — | 14.7 | |
| LSSModel=LSS2025.09 | — | — | — | 17.5 | |
| NegCLIPModel=NegCLIP2025.09 | — | — | — | 16.4 | |
| SigLIPModel=SigLIP2025.09 | — | — | — | 18.6 | |
| TinyLLaVALLM=Qwen1.5, Size=4B2026.07 | — | — | — | 70.1 | |
| TinyLLaVALLM=Qwen1.5, Size=1.8B2026.07 | — | — | — | 62.6 | |
| TinyLLaVALLM=Qwen1.5, Size=0.5B2026.07 | — | — | — | 50.2 | |
| TinyLLaVALLM=Qwen2.5, Size=3B2026.07 | — | — | — | 75 | |
| TinyLLaVALLM=Qwen2.5, Size=1.5B2026.07 | — | — | — | 70.3 | |
| TinyLLaVALLM=Qwen2.5, Size=0.5B2026.07 | — | — | — | 57 | |
| TRAGLLM=Qwen1.5, Size=1.8B2026.07 | — | — | — | 67.2 | |
| TRAGLLM=Qwen1.5, Size=0.5B2026.07 | — | — | — | 57.3 | |
| TRAGLLM=Qwen2.5, Size=1.5B2026.07 | — | — | — | 73.9 | |
| TRAGLLM=Qwen2.5, Size=0.5B2026.07 | — | — | — | 59.6 |