Multi-modal Evaluation on MME (test)
1,666Perception ScoreMini-Gemini
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Mini-GeminiLLM=Hermes-2-Yi-34B, Resolution=336, Zero-shot=true2024.03 | 1,666 | 439 | — | — | — | — | — | |
| Mini-Gemini-HDLLM=Hermes-2-Yi-34B, Resolution=672, Zero-shot=true2024.03 | 1,659 | 482 | — | — | — | — | — | |
| Mini-GeminiLLM=Mixtral-8x7B, Resolution=336, Zero-shot=true2024.03 | 1,639 | 379 | — | — | — | — | — | |
| Mini-Gemini-HDLLM=Mixtral-8x7B, Resolution=672, Zero-shot=true2024.03 | 1,633 | 356 | — | — | — | — | — | |
| LLaVA-NeXTLLM=Hermes-2-Yi-34B, Resolution=672, Zero-shot=true2024.03 | 1,631 | 397 | — | — | — | — | — | |
| Mini-Gemini-HDLLM=Vicuna-13B, Resolution=672, Zero-shot=true2024.03 | 1,597 | 320 | — | — | — | — | — | |
| LLaVA-NeXTLLM=Vicuna-13B, Resolution=672, Zero-shot=true2024.03 | 1,575 | 326 | — | — | — | — | — | |
| Mini-GeminiLLM=Vicuna-13B, Resolution=336, Zero-shot=true2024.03 | 1,565 | 322 | — | — | — | — | — | |
| Mini-Gemini-HDLLM=Vicuna-7B, Resolution=672, Zero-shot=true2024.03 | 1,546 | 319 | — | — | — | — | — | |
| LLaMA-VIDLLM=Vicuna-13B, Resolution=336, Zero-shot=true2024.03 | 1,542 | — | — | — | — | — | — | |
| LLaVA-1.5LLM=Vicuna-13B, Resolution=336, Zero-shot=true2024.03 | 1,531 | 295 | — | — | — | — | — | |
| Mini-GeminiLLM=Vicuna-7B, Resolution=336, Zero-shot=true2024.03 | 1,523 | 316 | — | — | — | — | — | |
| LLaMA-VIDLLM=Vicuna-7B, Resolution=336, Zero-shot=true2024.03 | 1,521 | — | — | — | — | — | — | |
| LLaVA-NeXTLLM=Vicuna-7B, Resolution=672, Zero-shot=true2024.03 | 1,519 | 332 | — | — | — | — | — | |
| ScalSelectSelection Budget=300K2026.02 | 1,517.26 | 290 | — | — | — | — | — | |
| LLaVA-1.5LLM=Vicuna-7B, Resolution=336, Zero-shot=true2024.03 | 1,511 | — | — | — | — | — | — | |
| Qwen-VL-ChatLLM=Qwen-7B, Resolution=448, Zero-shot=true2024.03 | 1,488 | — | — | — | — | — | — | |
| Full-FinetuneSelection Budget=625K (Full)2026.02 | 1,463.87 | 278.57 | — | — | — | — | — | |
| ScalSelectSelection Budget=200K2026.02 | 1,435.79 | 298.93 | — | — | — | — | — | |
| ScalSelectSelection Budget=400K2026.02 | 1,425.41 | 299.29 | — | — | — | — | — | |
| COINCIDESelection Budget=100K2026.02 | 1,422.01 | 274.29 | — | — | — | — | — | |
| RandomSelection Budget=100K2026.02 | 1,418.85 | 295.36 | — | — | — | — | — | |
| RDS+Selection Budget=100K2026.02 | 1,414.66 | 302.85 | — | — | — | — | — | |
| PerplexitySelection Budget=100K2026.02 | 1,403.08 | 278.93 | — | — | — | — | — | |
| ScalSelectSelection Budget=100K2026.02 | 1,400.34 | 308.57 | — | — | — | — | — | |
| PRISMSelection Budget=100K2026.02 | 1,396.24 | 276.43 | — | — | — | — | — | |
| Mini-GeminiLLM=Gemma-2B, Resolution=336, Zero-shot=true2024.03 | 1,341 | 312 | — | — | — | — | — | |
| OtterHDLLM=Fuyu-8B, Resolution=1024, Zero-shot=true2024.03 | 1,314 | — | — | — | — | — | — | |
| MobileVLMLLM=MiLAMA 2.7B, Resolution=336, Zero-shot=true2024.03 | 1,289 | — | — | — | — | — | — | |
| LengthSelection Budget=100K2026.02 | 1,282.97 | 292.14 | — | — | — | — | — | |
| ScalSelectSelection Budget=50K2026.02 | 1,282.22 | 314.29 | — | — | — | — | — | |
| InstructBLIPLLM=Vicuna-13B, Resolution=224, Zero-shot=true2024.03 | 1,213 | — | — | — | — | — | — | |
| DARTBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | — | — | 1,916 | — | — | 73.6 | 0.715 | |
| DART (ViT)Backbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | — | — | 2,198 | — | — | 88.1 | 0.639 | |
| FastVBase Model=LLaVA1.5-7B, Pruning Family=LLM Single-Layer Pruning2026.02 | — | — | 1,255 | 0.674 | — | — | — | |
| FastVBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | — | — | 1,977 | — | — | 78.6 | 0.707 | |
| InstructBLIPNumber of Parameters=13B2024.05 | — | — | 1,212.8 | — | — | — | — | |
| IPCV+FastVBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | — | — | 2,364 | — | — | 89.9 | 0.644 | |
| IVM-Enhanced LLaVA-7BNumber of Parameters=14B2024.05 | — | — | 1,806 | — | — | — | — | |
| LISA-Enhanced LLaVA-7BNumber of Parameters=20B2024.05 | — | — | 1,560 | — | — | — | — | |
| LLaVA-13BNumber of Parameters=13B2024.05 | — | — | 1,834 | — | — | — | — | |
| LLaVA-7BNumber of Parameters=7B2024.05 | — | — | 1,748 | — | — | — | — | |
| PDropBase Model=LLaVA1.5-7B, Pruning Family=LLM Multi-Layer Pruning2026.02 | — | — | 1,092 | 0.586 | — | — | — | |
| PoolingBase Model=LLaVA1.5-7B, Pruning Family=LLM Single-Layer Pruning2026.02 | — | — | 1,380 | 0.741 | — | — | — | |
| PoolingBase Model=LLaVA1.5-7B, Pruning Family=Vision Encoder-Side Pruning2026.02 | — | — | 1,151 | 0.618 | — | — | — | |
| PruMerge+Base Model=LLaVA1.5-7B, Pruning Family=Vision Encoder-Side Pruning2026.02 | — | — | 1,616 | 0.868 | — | — | — | |
| RandomBase Model=LLaVA1.5-7B, Pruning Family=LLM Single-Layer Pruning2026.02 | — | — | 1,351 | 0.726 | — | — | — | |
| RandomBase Model=LLaVA1.5-7B, Pruning Family=LLM Multi-Layer Pruning2026.02 | — | — | 1,342 | 0.721 | — | — | — | |
| RandomBase Model=LLaVA1.5-7B, Pruning Family=Vision Encoder-Side Pruning2026.02 | — | — | 1,410 | 0.757 | — | — | — | |
| SEAL-7BNumber of Parameters=7B2024.05 | — | — | 1,129 | — | — | — | — | |
| SparseVLMBase Model=LLaVA1.5-7B, Pruning Family=LLM Multi-Layer Pruning2026.02 | — | — | 1,589 | 0.853 | — | — | — | |
| SparseVLMBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | — | — | 1,916 | — | — | 75.1 | 0.734 | |
| ToFuBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | — | — | 2,074 | — | — | 83.4 | 0.619 | |
| ToMeBackbone=InternVL3-38B, ViT token retention rate=40%, LLM token retention rate=40%2025.12 | — | — | 2,092 | — | — | 82.7 | 0.617 | |
| V^2DropBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=5%2025.12 | — | — | 1,697 | — | — | 57.1 | 0.705 | |
| VanillaBase Model=LLaVA1.5-7B, Pruning Family=Baseline2026.02 | — | — | 1,862 | 1 | — | — | — | |
| VanillaBackbone=InternVL3-38B, ViT token retention rate=100%, LLM token retention rate=100%2025.12 | — | — | 2,467 | — | — | 100 | 1 | |
| VisionZipBase Model=LLaVA1.5-7B, Pruning Family=Vision Encoder-Side Pruning2026.02 | — | — | 1,690 | 0.908 | — | — | — |