Image-Text Retrieval on NaturalBench Retrieval
71.9T ScoreQwen2-VL + HEC-V
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2-VL + HEC-VBackbone=Qwen2-VL, Adaptation Method=HEC-V, Evaluation Protocol=2-way 20-shot2026.03 | 71.9 | 73 | 51.9 | |
| Qwen2-VL + SAVsBackbone=Qwen2-VL, Adaptation Method=SAVs, Evaluation Protocol=2-way 20-shot2026.03 | 70 | 71 | 42.5 | |
| Qwen2-VL + LoRABackbone=Qwen2-VL, Adaptation Method=LoRA finetuning, Evaluation Protocol=2-way 20-shot2026.03 | 65.2 | 66.1 | 40.4 | |
| GPT-4oEvaluation Mode=Zero-shot2026.03 | 65 | 67 | 40.5 | |
| Qwen2-VL + MTVBackbone=Qwen2-VL, Adaptation Method=MTV, Evaluation Protocol=2-way 20-shot2026.03 | 63.5 | 64 | 37 | |
| Qwen2-VLEvaluation Mode=Zero-shot2026.03 | 60.2 | 61.9 | 35.6 | |
| SigLipEvaluation Mode=Zero-shot2026.03 | 54.5 | 54.9 | 31.2 | |
| Qwen2-VL + 4-shot-ICLBackbone=Qwen2-VL, Evaluation Mode=4-shot-ICL2026.03 | 42.4 | 45.6 | 22.7 | |
| CLIPEvaluation Mode=Zero-shot2026.03 | 41.8 | 45 | 23.2 | |
| LLaVA-1.5Evaluation Mode=Zero-shot2026.03 | 36.7 | 42.7 | 12.2 | |
| Instruct-BLIPEvaluation Mode=Zero-shot2026.03 | 19.5 | 21.3 | 1.1 |