Knowledge-based VQA on InfoSeek
42.49Unseen-Q PerformanceMG2-RAG
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| MG2-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=Qwen3.5-27B, Retrieval Pool Size=5k2026.04 | 42.49 | 42.03 | 42.26 | |
| MG2-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=GPT-5.22026.04 | 40.16 | 38.47 | 39.3 | |
| MG2-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=Qwen2.5-VL-7B, Retrieval Pool Size=5k2026.04 | 39.17 | 38.15 | 38.65 | |
| MG2-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=Qwen3.5-27B2026.04 | 37.36 | 38.4 | 37.87 | |
| MG2-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=Qwen2.5-VL-7B2026.04 | 35.78 | 35.18 | 35.48 | |
| mKG-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=LLaMA-3.1-8B2026.04 | 32.9 | 31.3 | 32.1 | |
| GPT-5.2Model Category=Zero-Shot MLLMs, LLM/MLLM=GPT-5.22026.04 | 32.68 | 26.54 | 29.29 | |
| MG2-RAGModel Category=Graph Retrieval-Augmented Models, LLM/MLLM=LLaMA-3.1-8B2026.04 | 32.32 | 32.84 | 32.58 | |
| EchoSightModel Category=Retrieval-Augmented Models, LLM/MLLM=LLaMA-3.1-8B, Modality=V→T2026.04 | 30 | 30.7 | 30.4 | |
| Gemini-3.1-ProModel Category=Zero-Shot MLLMs, LLM/MLLM=Gemini-3.1-Pro2026.04 | 29.92 | 22.52 | 25.7 | |
| RORA-VLMModel Category=Retrieval-Augmented Models, LLM/MLLM=Vicuna-7B2026.04 | 25.1 | 27.3 | — | |
| Qwen2.5-VL-7BModel Category=Zero-Shot MLLMs, LLM/MLLM=Qwen2.5-VL-7B2026.04 | 22.8 | 24.1 | 23.7 | |
| Qwen3.5-27BModel Category=Zero-Shot MLLMs, LLM/MLLM=Qwen3.5-27B2026.04 | 20.25 | 16.75 | 18.33 | |
| EchoSightModel Category=Retrieval-Augmented Models, LLM/MLLM=LLaMA-3.1-8B, Modality=V→V2026.04 | 18 | 19.8 | 18.8 | |
| InstructBLIPModel Category=Zero-Shot MLLMs, LLM/MLLM=Flan-T5XL2026.04 | 8.9 | 7.4 | 8.1 | |
| LLaVA-v1.5Model Category=Zero-Shot MLLMs, LLM/MLLM=LLaMA-3.1-8B2026.04 | 8.3 | 8.9 | 7.8 | |
| VaLiK (5k)Model Category=Graph Retrieval-Augmented Models, LLM/MLLM=Qwen2.5-VL-7B, Retrieval Pool Size=5k2026.04 | 3.19 | 2.07 | 2.51 | |
| MMGraphRAG (5k)Model Category=Graph Retrieval-Augmented Models, LLM/MLLM=Qwen2.5-VL-7B, Retrieval Pool Size=5k2026.04 | 0.69 | 0.39 | 0.5 |