List your company
ResearchFormerly WizWand

Vision-Language Understanding on MMBench

88.7AccuracyQwen3-VL-4B-Instruct
-3.54820.40144.3568.299Mar 11, 2025May 29, 2025Aug 17, 2025Nov 5, 2025Jan 23, 2026Apr 13, 2026Jul 2, 2026
Updated 3mo ago

Evaluation results

MethodLinks
Qwen3-VL-4B-Instruct2026.0288.7
Qwen2.5-VL-72B-InstructAttack=Benign2026.0788.39
InternVL3.5-38BAttack=Benign2026.0786.99
Optimization-based Token SelectionTarget Model=InternVL3.5-38B2026.0786.22
Xiaomi-Robotics-0pre-training=with VL data2026.0284.4
InternVL3.5-8BAttack=Benign2026.0783.2
InternVL3.5-4BAttack=Benign2026.0781.03
MolmoAct2026.0280.1
Optimization-based Token SelectionTarget Model=InternVL3.5-8B2026.0780.01
Qwen2.5-VL-7B-InstructAttack=Benign2026.0779.87
Qwen2.5-VL-3B-InstructAttack=Benign2026.0776.54
SmoothSMoE annealed (k=2)k=2, Smoothing mechanism=Annealed SmoothSMoE, Backbone=ViT-based, Model parameters=5.6B, Fine-tuning dataset=50% of the LLaVA-665K dataset, Number of experts=42026.0671.13
SMoE (k=2)k=2, Smoothing mechanism=None, Backbone=ViT-based, Model parameters=5.6B, Fine-tuning dataset=50% of the LLaVA-665K dataset, Number of experts=42026.0670.96
SmoothSMoE (k=2.5)k=2.5, Smoothing mechanism=SmoothSMoE, Backbone=ViT-based, Model parameters=5.6B, Fine-tuning dataset=50% of the LLaVA-665K dataset, Number of experts=42026.0669.42
Upper BoundBackbone=LLaVA-v1.5-13B, Retained Tokens=Full, Pruning Ratio=0%2026.0468.9
CLASPBackbone=LLaVA-v1.5-13B, Retained Tokens=192, Pruning Ratio=66.7%2026.0468
CLASPBackbone=LLaVA-v1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.0468
VanillaBackbone=LLaVA-NeXT-7B, Token Budget=2880 Tokens, Venue=-2026.0267.4
VanillaBase Model=LLaVA-Next-7B, Token Reduction Rate=0.0%, N=28802025.0567.4
SparseVLMBackbone=LLaVA-v1.5-13B, Retained Tokens=192, Pruning Ratio=66.7%2026.0467.4
VanillaBackbone=LLaVA-NeXT-7B, Retained Tokens=2880, Compression Ratio=100%2026.0467.4
PIO-FVLMBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=Ours2026.0266.2
MoBBase Model=LLaVA-Next-7B, Objectives=PA VP, Pruning budget K=320, Token Reduction Rate=88.9%, η-prior=true2025.0565.8
SparseVLMBackbone=LLaVA-v1.5-13B, Retained Tokens=128, Pruning Ratio=77.8%2026.0465.8
CDPrunerBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=NeurIPS’252026.0265.5
DARTBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=EMNLP’252026.0265.3
HoloVBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=NeurIPS’252026.0265.3
VanillaBase Model=LLaVA-1.5-7B, Token Reduction Rate=0.0%, N=5762025.0564.7
CLASPBackbone=LLaVA-v1.5-13B, Retained Tokens=64, Pruning Ratio=88.9%2026.0464.4
MoBBase Model=LLaVA-1.5-7B, Pruning budget K=192, Token Reduction Rate=66.7%, η-prior=true2025.0564.1
TwigVLMBase Model=LLaVA-1.5-7B, Objectives=PA, Pruning budget K=192, Token Reduction Rate=66.7%2025.0564
MoBBase Model=LLaVA-1.5-7B, Objectives=PA VP, Pruning budget K=192, Token Reduction Rate=66.7%, η-prior=false2025.0563.8
DARTBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0563.6
MoBBase Model=LLaVA-1.5-7B, Pruning budget K=128, Token Reduction Rate=77.8%, η-prior=true2025.0563.5
PyramidDropBase Model=LLaVA-1.5-7B, Objectives=PA, Pruning budget K=192, Token Reduction Rate=66.7%2025.0563.3
VisionZipBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=CVPR’252026.0263.1
VisionZipBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0563
TokenCarveBase Model=LLaVA-1.5-7B, Objectives=PA VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0563
HiREDBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0562.8
EvoCompBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%, l=02026.0462.7
SparseVLMBase Model=LLaVA-1.5-7B, Objectives=PA, Pruning budget K=192, Token Reduction Rate=66.7%2025.0562.5
FiCoCo-VBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0562.3
MustDropBase Model=LLaVA-1.5-7B, Objectives=PA VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0562.3
MoBBase Model=LLaVA-1.5-7B, Pruning budget K=64, Token Reduction Rate=88.9%, η-prior=true2025.0562.1
MustDropBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0461.7
FastVBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=ECCV’242026.0261.6
SparseVLMBackbone=LLaVA-v1.5-13B, Retained Tokens=64, Pruning Ratio=88.9%2026.0461.3
FastVBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0561.2
DARTBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0461.2
SparseVLMBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0461.1
CLASPBackbone=LLaVA-v1.5-13B, Retained Tokens=32, Pruning Ratio=94.4%2026.0460.7
SparseVLMBackbone=LLaVA-NeXT-7B, Token Budget=320 Tokens, Venue=ICML’252026.0260.6
ToMeBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0560.5
FastVBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0460.3
VisionZipBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0460.1
PyramidDropBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0460
LLaVA-PruMergeBase Model=LLaVA-1.5-7B, Objectives=VP, Pruning budget K=192, Token Reduction Rate=66.7%2025.0559.6
GlobalCom^2Backbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%2026.0459.1
MolmoE-1B-7BNA=1.58B, Avgk=8.002026.0658.24
AdaMoENA=1.39B, Avgk=6.132026.0656.95
GeMoENA=1.32B, Avgk=5.432026.0656.7
SparseVLMBackbone=LLaVA-v1.5-13B, Retained Tokens=32, Pruning Ratio=94.4%2026.0456.2
DARTBackbone=LLaVA-NeXT-7B, Retained Tokens=160, Compression Ratio=94.4%, l=02026.0456.2
Top-pNA=2.05B, Avgk=12.652026.0655.75
MMaDA+VRCDL=192, FR=0.252026.0554.84
MMaDA+VRCDL=384, FR=0.252026.0553.52
MMaDAL=192, FR=0.252026.0553.45
MMaDA+VRCDL=192, FR=0.1252026.0553.45
MMaDA+VRCDL=192, FR=0.52026.0552.87
MMaDA+VRCDL=384, FR=0.1252026.0552.81
MMaDAL=192, FR=0.1252026.0552.09
GKL-KDCLIP Model=GKL-KD2025.0352.06
KL-KDCLIP Model=KL-KD [27]2025.0351.89
DYNMoENA=1.40B, Avgk=6.192026.0651.72
MMaDA+VRCDL=384, FR=0.52026.0551.32
MoE++NA=1.48B, Avgk=7.032026.0651.2
MMaDAL=384, FR=0.1252026.0550.9
MMaDAL=384, FR=0.252026.0550.82
MMaDAL=192, FR=0.52026.0550.66
BaselineCLIP Model=Baseline2025.0349.66
MMaDAL=384, FR=0.52026.0544.65
Optimization-based Token SelectionTarget Model=InternVL3.5-4B2026.0725
pi_0.52026.0222.1
Optimization-based Token SelectionTarget Model=Qwen2.5-VL-72B-Instruct2026.070.08
pi_02026.020
Xiaomi-Robotics-0pre-training=without VL data2026.020
Optimization-based Token SelectionTarget Model=Qwen2.5-VL-3B-Instruct2026.070
Optimization-based Token SelectionTarget Model=Qwen2.5-VL-7B-Instruct2026.070