Knowledge-Based Visual Question Answering on ViQuAE (test)
61Overall AccuracyInternVL3.5-38B + LoT
Evaluation Results
| Method | Links | |
|---|---|---|
| InternVL3.5-38B + LoTBackbone=InternVL3.5-38B, Model Scale=Large, Method=LoT2026.04 | 61 | |
| InternVL3.5-8B + LoTBackbone=InternVL3.5-8B, Model Scale=Medium, Method=LoT2026.04 | 54.2 | |
| InternVL3.5-38BBackbone=InternVL3.5-38B, Model Scale=Large, Method=Zero-shot retrieval-augmented2026.04 | 51.5 | |
| Qwen3-VL-8B + LoTBackbone=Qwen3-VL-8B, Model Scale=Medium, Method=LoT2026.04 | 51 | |
| Qwen2.5-VL-32B + LoTBackbone=Qwen2.5-VL-32B, Model Scale=Large, Method=LoT2026.04 | 46 | |
| InternVL3.5-4B + LoTBackbone=InternVL3.5-4B, Model Scale=Small, Method=LoT2026.04 | 45.6 | |
| InternVL3.5-8BBackbone=InternVL3.5-8B, Model Scale=Medium, Method=Zero-shot retrieval-augmented2026.04 | 44.2 | |
| Qwen3-VL-8BBackbone=Qwen3-VL-8B, Model Scale=Medium, Method=Zero-shot retrieval-augmented2026.04 | 43.7 | |
| Qwen2-VL-7B + LoTBackbone=Qwen2-VL-7B, Model Scale=Medium, Method=LoT2026.04 | 40.8 | |
| Qwen2.5-VL-7B + LoTBackbone=Qwen2.5-VL-7B, Model Scale=Medium, Method=LoT2026.04 | 38.6 | |
| Qwen2.5-VL-32BBackbone=Qwen2.5-VL-32B, Model Scale=Large, Method=Zero-shot retrieval-augmented2026.04 | 37.9 | |
| InternVL3.5-4BBackbone=InternVL3.5-4B, Model Scale=Small, Method=Zero-shot retrieval-augmented2026.04 | 36.4 | |
| Qwen3-VL-4B + LoTBackbone=Qwen3-VL-4B, Model Scale=Small, Method=LoT2026.04 | 36.3 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B, Model Scale=Medium, Method=Zero-shot retrieval-augmented2026.04 | 36.3 | |
| Qwen3-VL-4BBackbone=Qwen3-VL-4B, Model Scale=Small, Method=Zero-shot retrieval-augmented2026.04 | 34.5 | |
| Qwen2-VL-7BBackbone=Qwen2-VL-7B, Model Scale=Medium, Method=Zero-shot retrieval-augmented2026.04 | 33 | |
| Qwen2.5-VL-3B + LoTBackbone=Qwen2.5-VL-3B, Model Scale=Small, Method=LoT2026.04 | 27.8 | |
| Qwen2.5-VL-3BBackbone=Qwen2.5-VL-3B, Model Scale=Small, Method=Zero-shot retrieval-augmented2026.04 | 22.9 | |
| Qwen2-VL-2B + LoTBackbone=Qwen2-VL-2B, Model Scale=Small, Method=LoT2026.04 | 18.4 | |
| Qwen2-VL-2BBackbone=Qwen2-VL-2B, Model Scale=Small, Method=Zero-shot retrieval-augmented2026.04 | 15 |