List your company
ResearchFormerly WizWand

Visual Perception on MME Perception

1,742MME^PInternVL2-7B + SCR
1,081.081,252.6651,424.251,595.835Feb 7, 2026Feb 12, 2026Feb 18, 2026Feb 23, 2026Mar 1, 2026Mar 6, 2026Mar 12, 2026
Updated 7mo ago

Evaluation results

MethodLinks
InternVL2-7B + SCRConfiguration=Spatial Credit Redistribution2026.021,742
InternVL2-7BConfiguration=Vanilla2026.021,685
Qwen2-VL-7B + SCRConfiguration=Spatial Credit Redistribution2026.021,678
LLaVA-13B + SCRConfiguration=Spatial Credit Redistribution2026.021,642
Qwen2-VL-7BConfiguration=Vanilla2026.021,620
Qwen-VL + SCRConfiguration=Spatial Credit Redistribution2026.021,612
LLaVA-13BConfiguration=Vanilla2026.021,588
LLaVA-7B + SCRConfiguration=Spatial Credit Redistribution2026.021,576
ShareGPT4VLLM=Vicuna-7B, Resolution=3362026.031,567.4
PDrop (training-based)Base Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=270, Vision-side TFLOPs (Rel.)=6.92 (46.4%)2026.021,555.2
Dynamic-LLaVABase Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=115, Vision-side TFLOPs (Rel.)=2.93 (19.7%)2026.021,554.1
Qwen-VLConfiguration=Vanilla2026.021,548
LLaVA-v1.5LLM=Vicuna-1.5-13B, Resolution=3362026.031,531.3
Original (LLaVA-1.5-13B)Base Model=LLaVA-1.5-13B, Pruning Strategy=None (Baseline), Sparsity (Op.)=false, Sparsity (Tok.)=false, Vision-side Token=576, Vision-side TFLOPs (Rel.)=14.91 (100.0%)2026.021,529.9
Delta-LLaVABase Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=144, Vision-side TFLOPs (Rel.)=3.68 (24.7%)2026.021,527.5
LLaVA-7BConfiguration=Vanilla2026.021,510
TwigVLMBase Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=64, Vision-side TFLOPs (Rel.)=1.63 (10.9%)2026.021,503.1
TokLIPLLM=Qwen2.5-7B-Inst., Resolution=3842026.031,488.4
Qwen-VL-ChatLLM=Qwen-7B, Resolution=4482026.031,478.5
DOP_CDBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=true, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,468.9
VISABase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=64, Vision-side TFLOPs (Rel.)=1.63 (10.9%)2026.021,468.3
EvoTokLLM=Qwen-2.5-7B-Inst., Resolution=2562026.031,455.9
ViCABase Model=LLaVA-1.5-13B, Pruning Strategy=training-based (our architecture), Sparsity (Op.)=true, Sparsity (Tok.)=false, Vision-side Token=19, Vision-side TFLOPs (Rel.)=0.49 (3.3%)2026.021,453.6
ILLUMELLM=Vicuna-7B, Resolution=-2026.031,445.3
ViCA+PDrop†Base Model=LLaVA-1.5-13B, Pruning Strategy=training-based (our architecture + inference drop), Sparsity (Op.)=true, Sparsity (Tok.)=true, Vision-side Token=10, Vision-side TFLOPs (Rel.)=0.26 (1.7%)2026.021,442.1
YOPOBase Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=70, Vision-side TFLOPs (Rel.)=1.78 (12.0%)2026.021,430
LLaVA-PruMergeBase Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,428.6
CDPrunerBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,421
Cham-30B + SCRConfiguration=Spatial Credit Redistribution2026.021,418
DivPruneBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,405.2
SparseVLMBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=64, Vision-side TFLOPs (Rel.)=1.63 (10.9%)2026.021,374.3
DOP_vBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=true, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,365.6
TokenFlow-LLLM=Vicuna-13B, Resolution=2562026.031,365.4
Cham-7B + SCRConfiguration=Spatial Credit Redistribution2026.021,358
SemHiTokLLM=Qwen2.5-7B, Resolution=2562026.031,355.8
TokenFlow-BLLM=Vicuna-13B, Resolution=2242026.031,353.6
Cham-30BConfiguration=Vanilla2026.021,342
TRIMBase Model=LLaVA-1.5-13B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=29, Vision-side TFLOPs (Rel.)=0.74 (4.9%)2026.021,337.9
VisPrunerBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,314.2
ToMeBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=64, Vision-side TFLOPs (Rel.)=1.63 (10.9%)2026.021,287
Cham-7BConfiguration=Vanilla2026.021,285
DARTBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,282.8
Original (LLaVA-1.5-3B)Base Model=LLaVA-1.5-3B, Pruning Strategy=None (Baseline), Sparsity (Op.)=false, Sparsity (Tok.)=false, Vision-side Token=576, Vision-side TFLOPs (Rel.)=3.04 (100.0%)2026.021,258.2
VisionZip (training-free)Base Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=32, Vision-side TFLOPs (Rel.)=0.81 (5.5%)2026.021,257.7
PDrop (training-free)Base Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=64, Vision-side TFLOPs (Rel.)=1.63 (10.9%)2026.021,247
FastVBase Model=LLaVA-1.5-13B, Pruning Strategy=training-free, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=64, Vision-side TFLOPs (Rel.)=1.63 (10.9%)2026.021,246.4
PDropBase Model=LLaVA-1.5-3B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=270, Vision-side TFLOPs (Rel.)=1.40 (46.0%)2026.021,231.1
ViCABase Model=LLaVA-1.5-3B, Pruning Strategy=training-based (our architecture), Sparsity (Op.)=true, Sparsity (Tok.)=false, Vision-side Token=27, Vision-side TFLOPs (Rel.)=0.14 (4.5%)2026.021,188.9
ViCA+PDrop†Base Model=LLaVA-1.5-3B, Pruning Strategy=training-based (our architecture + inference drop), Sparsity (Op.)=true, Sparsity (Tok.)=true, Vision-side Token=12, Vision-side TFLOPs (Rel.)=0.06 (2.0%)2026.021,187.5
YOPOBase Model=LLaVA-1.5-3B, Pruning Strategy=training-based, Sparsity (Op.)=false, Sparsity (Tok.)=true, Vision-side Token=70, Vision-side TFLOPs (Rel.)=0.37 (12.0%)2026.021,106.5