Multimodal Question Answering on MMBench en (test)
89AccuracyVanilla
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VanillaBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=100%2025.12 | 89 | 100 | 100 | |
| DART (ViT)Backbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 85 | 88.1 | 63.9 | |
| InternVL2.5-8B2026.04 | 84.6 | — | — | |
| IPCV+FastVBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 84.5 | 89.9 | 64.4 | |
| ToFuBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 83.9 | 83.4 | 61.9 | |
| GPT-4o2026.04 | 83.4 | — | — | |
| Qwen2.5-VL-7BTraining Stage=Base2026.04 | 82.8 | — | — | |
| Claude-3.5 Sonnet2026.04 | 82.6 | — | — | |
| Insight-V-8B2026.04 | 82.3 | — | — | |
| ToMeBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | 82.2 | 82.7 | 61.7 | |
| IXC-2.52026.04 | 82.2 | — | — | |
| Qwen2.5-VL-7B + Saliency-R1Training Stage=Saliency-R12026.04 | 81.8 | — | — | |
| InternVL2-8B2026.04 | 81.7 | — | — | |
| Qwen2.5-VL-7B + SFTTraining Stage=SFT2026.04 | 78.8 | — | — | |
| MiniCPM-V-2.6-8B2026.04 | 78 | — | — | |
| Qwen2.5-VL-3BTraining Stage=Base2026.04 | 77.8 | — | — | |
| MiniCPM-Llama-V-2.5-8B2026.04 | 77.2 | — | — | |
| Qwen2.5-VL-3B + Saliency-R1Training Stage=Saliency-R12026.04 | 76.7 | — | — | |
| FastVBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 76.6 | 78.6 | 70.7 | |
| Cambrain-1-8B2026.04 | 75.9 | — | — | |
| Qwen2.5-VL-3B + SFTTraining Stage=SFT2026.04 | 75.8 | — | — | |
| LLaVA-CoT-11B2026.04 | 75 | — | — | |
| SparseVLMBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 74.6 | 75.1 | 73.4 | |
| Gemini-1.5-Pro2026.04 | 73.9 | — | — | |
| DARTBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 71.1 | 73.6 | 71.5 | |
| V^2DropBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | 59.5 | 57.1 | 70.5 |