Multimodal Evaluation on MME (MME-P and MME-C)
1,864MME-P ScoreTwigVLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TwigVLMRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.03 | 1,864 | — | |
| TwigVLM++Retained Tokens=640, Base Model=LLaVA-NeXT-7B2025.03 | 1,858 | — | |
| LLaVA-NeXT-7BRetained Tokens=2880, Base Model=LLaVA-NeXT-7B2025.03 | 1,851 | — | |
| FastVRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.03 | 1,807 | — | |
| VisionZipRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.03 | 1,787 | — | |
| VisionZip‡Retained Tokens=640, Base Model=LLaVA-NeXT-7B2025.03 | 1,778 | — | |
| TwigVLM++Retained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,773 | — | |
| SparseVLMRetained Tokens=640, Base Model=LLaVA-NeXT-7B2025.03 | 1,772 | — | |
| VisionZip‡Retained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,770 | — | |
| TwigVLMRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,758 | — | |
| VisionZipRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,702 | — | |
| FP16 (Baseline)Backbone=Qwen 2.5 VL 7B, Avg. Bit width=162025.09 | 1,695 | 640 | |
| SparseVLMRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,694 | — | |
| OmniQuantBackbone=Qwen 2.5 VL 7B, Avg. Bit width=42025.09 | 1,655 | 610 | |
| Qwen2.5VL 7B# Vis tok.=14002025.04 | 1,652.7 | 633.4 | |
| AWQBackbone=Qwen 2.5 VL 7B, Avg. Bit width=42025.09 | 1,645 | 620 | |
| MustDropRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,641 | — | |
| LUQBackbone=Qwen 2.5 VL 7B, Avg. Bit width=2.75, Quantized Layers=12 layer2025.09 | 1,640 | 600 | |
| FLARE-X 8B# Vis tok.=14002025.04 | 1,639.1 | 378.9 | |
| GPTQBackbone=Qwen 2.5 VL 7B, Avg. Bit width=42025.09 | 1,638 | 610 | |
| ETAModel=InternLM-XComposer-2.52025.10 | 1,629.4 | 546.1 | |
| LLaVA-OneVision 7B# Vis tok.=24002025.04 | 1,626 | 483 | |
| VanillaModel=InternLM-XComposer-2.52025.10 | 1,623.7 | 551.1 | |
| ECSOModel=InternLM-XComposer-2.52025.10 | 1,623.7 | 551.1 | |
| MoRASModel=InternLM-XComposer-2.52025.10 | 1,623.7 | 551.1 | |
| ASTRAModel=InternLM-XComposer-2.52025.10 | 1,617.4 | 546.8 | |
| FLARE-X 3B# Vis tok.=14002025.04 | 1,612.4 | 437.2 | |
| CoCAModel=InternLM-XComposer-2.52025.10 | 1,606.5 | 551.1 | |
| MiniCPM-Llama3-V-2.5 8B# Vis tok.=4002025.04 | 1,604.8 | 398.5 | |
| FLARE-L 8B# Vis tok.=6302025.04 | 1,596.8 | 347.3 | |
| Qwen2.5VL 3B# Vis tok.=14002025.04 | 1,592.4 | 607.5 | |
| AutoSteerModel=InternLM-XComposer-2.52025.10 | 1,591.2 | 544.3 | |
| FLARE-L 3B# Vis tok.=6302025.04 | 1,588.3 | 425.6 | |
| FigStepModel=InternLM-XComposer-2.52025.10 | 1,577.7 | 516.8 | |
| MoRASModel=LLaVA-OneVision-7B2025.10 | 1,561.8 | 409.6 | |
| VanillaModel=LLaVA-OneVision-7B2025.10 | 1,560.9 | 409.6 | |
| ECSOModel=LLaVA-OneVision-7B2025.10 | 1,560.9 | 409.6 | |
| ETAModel=LLaVA-OneVision-7B2025.10 | 1,560.9 | 409.6 | |
| Florence-VL 8B# Vis tok.=5762025.04 | 1,560 | 381.1 | |
| Eagle 8B# Vis tok.=10242025.04 | 1,559 | — | |
| DeepSeek-VL2-Tiny# Vis tok.=15002025.04 | 1,548.3 | 357.1 | |
| AutoSteerModel=LLaVA-OneVision-7B2025.10 | 1,548.1 | 409.6 | |
| Cambrain 8B# Vis tok.=5762025.04 | 1,547.1 | — | |
| FastVRetained Tokens=320, Base Model=LLaVA-NeXT-7B2025.03 | 1,539 | — | |
| Molmo 7B# Vis tok.=12002025.04 | 1,536.1 | 341.6 | |
| MiniCPM-V-2.0 3B# Vis tok.=4002025.04 | 1,531.4 | 322.8 | |
| ETABackbone=LLaVA-1.5-13B2025.10 | 1,531.2 | 296.1 | |
| VanillaBackbone=LLaVA-1.5-13B2025.10 | 1,529.9 | 298.6 | |
| ECSOBackbone=LLaVA-1.5-13B2025.10 | 1,529.9 | 298.6 | |
| MoRASBackbone=LLaVA-1.5-13B2025.10 | 1,529.9 | 298.6 | |
| CoCAModel=LLaVA-1.5-7B2025.10 | 1,526.5 | 283.6 | |
| FullData %=100%2026.05 | 1,517.3 | 656.1 | |
| GRACEData %=20%2026.05 | 1,512.3 | 682.9 | |
| LESSData %=20%2026.05 | 1,511.4 | 668.6 | |
| AutoSteerBackbone=LLaVA-1.5-13B2025.10 | 1,510.9 | 296.4 | |
| FP16 (Baseline)Backbone=LLaVA-1.5 7B, Avg. Bit width=162025.09 | 1,510 | 350 | |
| ETAModel=LLaVA-1.5-7B2025.10 | 1,509.3 | 339.6 | |
| GRACEData %=5%2026.05 | 1,509.3 | 617.5 | |
| LLaVA1.5-7BPruning budget (keep tokens)=Full (No pruning), Base model=LLaVA1.5-7B2026.04 | 1,509.13 | — | |
| VanillaModel=LLaVA-1.5-7B2025.10 | 1,505.1 | 355.7 | |
| ECSOModel=LLaVA-1.5-7B2025.10 | 1,505.1 | 355.7 | |
| MoRASModel=LLaVA-1.5-7B2025.10 | 1,505.1 | 355.7 | |
| GRACEData %=10%2026.05 | 1,500.7 | 660.4 | |
| GRACEData %=15%2026.05 | 1,500.5 | 679.3 | |
| CADCData %=20%2026.05 | 1,498.9 | 630.7 | |
| Florence-VL 3B# Vis tok.=5762025.04 | 1,498.7 | 403.9 | |
| ICONSData %=20%2026.05 | 1,497 | 675 | |
| RandomData %=20%2026.05 | 1,495.5 | 653.2 | |
| LongestData %=20%2026.05 | 1,494.5 | 687.5 | |
| VanillaModel=Qwen-VL-Chat2025.10 | 1,489.9 | 331.8 | |
| ECSOModel=Qwen-VL-Chat2025.10 | 1,489.9 | 331.8 | |
| MoRASModel=Qwen-VL-Chat2025.10 | 1,489.9 | 331.8 | |
| ETAModel=Qwen-VL-Chat2025.10 | 1,487.9 | 331.8 | |
| FigStepModel=Qwen-VL-Chat2025.10 | 1,480.9 | 296.4 | |
| AutoSteerModel=LLaVA-1.5-7B2025.10 | 1,479.9 | 315.4 | |
| StepmaxData %=20%2026.05 | 1,477.7 | 637.5 | |
| AutoSteerModel=Qwen-VL-Chat2025.10 | 1,474.5 | 348.2 | |
| Phi 3.5-Vision# Vis tok.=21002025.04 | 1,473.4 | 412.1 | |
| CoCABackbone=LLaVA-1.5-13B2025.10 | 1,472.8 | 301.8 | |
| FigStepModel=LLaVA-OneVision-7B2025.10 | 1,472.2 | 395.4 | |
| ASTRAModel=LLaVA-1.5-7B2025.10 | 1,472 | 331.4 | |
| OmniQuantBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.09 | 1,466 | 318 | |
| ASTRAModel=Qwen-VL-Chat2025.10 | 1,465.9 | 323.2 | |
| SlimLLMBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.09 | 1,465 | 345 | |
| CoCAModel=LLaVA-OneVision-7B2025.10 | 1,463.4 | 403.2 | |
| AWQBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.09 | 1,456 | 349 | |
| GPTQBackbone=LLaVA-1.5 7B, Avg. Bit width=42025.09 | 1,450 | 347 | |
| FigStepModel=LLaVA-1.5-7B2025.10 | 1,435.7 | 275 | |
| SlimLLMBackbone=LLaVA-1.5 7B, Avg. Bit width=32025.09 | 1,429 | 329 | |
| MI-AttentionPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.04 | 1,428.03 | — | |
| FigStepBackbone=LLaVA-1.5-13B2025.10 | 1,423.9 | 322.1 | |
| ASTRAModel=LLaVA-OneVision-7B2025.10 | 1,406.4 | 387.9 | |
| MI-PrunerPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.04 | 1,381.35 | — | |
| CoCAModel=Qwen-VL-Chat2025.10 | 1,377.1 | 319.3 | |
| ASTRABackbone=LLaVA-1.5-13B2025.10 | 1,371.8 | 327.5 | |
| VisPrunerPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.04 | 1,369.9 | — | |
| VisionZipPruning budget (keep tokens)=64, Base model=LLaVA1.5-7B2026.04 | 1,365.6 | — | |
| LUQBackbone=LLaVA-1.5 7B, Avg. Bit width=2.54, Quantized Layers=16 layer2025.09 | 1,365 | 257 | |
| GPTQBackbone=LLaVA-1.5 7B, Avg. Bit width=32025.09 | 1,346 | 273 | |
| MI-AttentionPruning budget (keep tokens)=32, Base model=LLaVA1.5-7B2026.04 | 1,344.16 | — |