Visual Retrieval-Augmented Generation on Visual-RAG
66.04ScoreCLIP-B
Evaluation Results
| Method | Links | |
|---|---|---|
| CLIP-BMain Model=Ovis2.5-9B, Top K=1, Retriever Params=151M2026.05 | 66.04 | |
| Ours (Ovis2.5-2B Surrogate)Main Model=MiniCPM-V4.5, Top K=3, Retriever Params=2.6B2026.05 | 65.11 | |
| VLM2Vec-V2.0Main Model=Qwen3-VL-8B, Top K=4, Retriever Params=2.2B2026.05 | 61.63 | |
| OpenCLIPMain Model=InternVL3.5-8B, Top K=1, Retriever Params=151M2026.05 | 60.29 | |
| GT (Human-Annotated Image as Input)Main Model=InternVL3.5-8B, Top K=52026.05 | 58.82 | |
| BGE-MLLM-S2Main Model=Gemma3-12B, Top K=2, Retriever Params=7.6B2026.05 | 55.75 | |
| ViLoMemBackbone=QWEN2.5-VL-32B, Augmentation=ViLoMem2026.01 | 54.6 | |
| PolarMemBackbone=QWEN2.5-VL-32B, Augmentation=PolarMem2026.01 | 53.4 | |
| Zero-ShotMain Model=Ovis2.5-9B2026.05 | 52.67 | |
| ViLoMemBackbone=QWEN2.5-VL-7B, Augmentation=ViLoMem2026.01 | 50.8 | |
| Vanilla RAGBackbone=QWEN2.5-VL-32B, Augmentation=Vanilla RAG2026.01 | 50 | |
| PolarMemBackbone=QWEN2.5-VL-7B, Augmentation=PolarMem2026.01 | 49.9 | |
| PolarMemBackbone=LLaVA-Next-Llama-3-8B, Augmentation=PolarMem2026.01 | 48.6 | |
| QWEN2.5-VL-32BBackbone=QWEN2.5-VL-32B, Augmentation=None2026.01 | 47.9 | |
| ViLoMemBackbone=LLaVA-Next-Llama-3-8B, Augmentation=ViLoMem2026.01 | 47.9 | |
| Vanilla RAGBackbone=QWEN2.5-VL-7B, Augmentation=Vanilla RAG2026.01 | 44.1 | |
| PolarMemBackbone=DeepSeek-VL2, Augmentation=PolarMem2026.01 | 44.1 | |
| PolarMemBackbone=InternVL-3.5-4B, Augmentation=PolarMem2026.01 | 43.6 | |
| ViLoMemBackbone=DeepSeek-VL2, Augmentation=ViLoMem2026.01 | 42.7 | |
| PolarMemBackbone=LLaVA-Next-Mistral-7B, Augmentation=PolarMem2026.01 | 42.6 | |
| ViLoMemBackbone=InternVL-3.5-8B, Augmentation=ViLoMem2026.01 | 42 | |
| ViLoMemBackbone=LLaVA-Next-Mistral-7B, Augmentation=ViLoMem2026.01 | 41.8 | |
| PolarMemBackbone=InternVL-3.5-8B, Augmentation=PolarMem2026.01 | 41.3 | |
| QWEN2.5-VL-7BBackbone=QWEN2.5-VL-7B, Augmentation=None2026.01 | 40.2 | |
| ViLoMemBackbone=DeepSeek-VL2-Small, Augmentation=ViLoMem2026.01 | 40.2 | |
| InternVL-3.5-8BBackbone=InternVL-3.5-8B, Augmentation=None2026.01 | 40.1 | |
| ViLoMemBackbone=InternVL-3.5-4B, Augmentation=ViLoMem2026.01 | 40 | |
| Vanilla RAGBackbone=InternVL-3.5-8B, Augmentation=Vanilla RAG2026.01 | 39.6 | |
| PolarMemBackbone=DeepSeek-VL2-Small, Augmentation=PolarMem2026.01 | 39.5 | |
| Vanilla RAGBackbone=DeepSeek-VL2, Augmentation=Vanilla RAG2026.01 | 39.4 | |
| DeepSeek-VL2Backbone=DeepSeek-VL2, Augmentation=None2026.01 | 38.2 | |
| Vanilla RAGBackbone=InternVL-3.5-4B, Augmentation=Vanilla RAG2026.01 | 38.1 | |
| Vanilla RAGBackbone=LLaVA-Next-Llama-3-8B, Augmentation=Vanilla RAG2026.01 | 37.9 | |
| InternVL-3.5-4BBackbone=InternVL-3.5-4B, Augmentation=None2026.01 | 37.8 | |
| LLaVA-Next-Llama-3-8BBackbone=LLaVA-Next-Llama-3-8B, Augmentation=None2026.01 | 36.8 | |
| Vanilla RAGBackbone=LLaVA-Next-Mistral-7B, Augmentation=Vanilla RAG2026.01 | 35.9 | |
| Vanilla RAGBackbone=DeepSeek-VL2-Small, Augmentation=Vanilla RAG2026.01 | 35.6 | |
| DeepSeek-VL2-SmallBackbone=DeepSeek-VL2-Small, Augmentation=None2026.01 | 35.2 | |
| LLaVA-Next-Mistral-7BBackbone=LLaVA-Next-Mistral-7B, Augmentation=None2026.01 | 35.2 |