List your company
ResearchFormerly WizWand

Image Understanding on MME

2,500ScoreVanilla MoE
1,083.521,451.261,8192,186.74Nov 19, 2025Dec 19, 2025Jan 18, 2026Feb 18, 2026Mar 20, 2026Apr 19, 2026May 20, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Vanilla MoEBackbone=QWEN3-VL-30B-A3B-INSTRUCT2026.042,500
MACS (Ours)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.042,492
MACS (w/o Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.042,478
BAGEL-Uni-Edit (Ours)Training Stage=Stage 22026.052,412
BAGEL-Uni-Edit (Ours)Training Stage=Stage 12026.052,405
BAGEL2026.052,381
BAGEL-RecA2026.052,381
Uni-Edit w/o joint trainingjoint training=false2026.052,381
Uni-Edit w/o VAE dropoutVAE dropout=false2026.052,344
Uni-Edit ViT (und.) 224x980Resolution=224x9802026.052,326
Vanilla MoEBackbone=INTERNVL3.5-30B-A3B2026.042,324
Uni-Edit ViT (und.) 224x224Resolution=224x2242026.052,324
MACS (Ours)Backbone=INTERNVL3.5-30B-A3B2026.042,315
BAGEL-AnyEdit2026.052,314
Top-k RoutingExpert skipping ratio (ρ)=0, Number of experts (k)=8, Backbone=InternVL-3.5-30B-A3B-HF2025.112,312
Uni-Edit ViT (und.) 378x980Resolution=378x9802026.052,307
CAI-MoE (Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.042,305
MACS (w/o Expanded)Backbone=INTERNVL3.5-30B-A3B2026.042,298
MoDESExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.112,289
MoDESExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.112,258
MoDESExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.112,222
Vanilla MoEBackbone=KIMI-VL-A3B-INSTRUCT2026.042,218
CAI-MoE (Token Drop)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.042,214
MACS (Ours)Backbone=KIMI-VL-A3B-INSTRUCT2026.042,212
Kimi-VL-A3B-Instructk=6, Status=Default2025.112,207
MoDESSkip x% Experts=67%, rho=0.652025.112,204
MoDESSkip x% Experts=50%, rho=0.482025.112,203
MACS (w/o Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.042,198
MC-MoEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.112,192
MoDESSkip x% Experts=83%, rho=0.82025.112,162
CAI-MoE (Expanded)Backbone=INTERNVL3.5-30B-A3B2026.042,147
CAI-MoE (Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.042,096
DiEPExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.112,084
+Uni-Edit (Ours)Base Model=Janus-Pro2026.052,067
CAI-MoE (Token Drop)Backbone=INTERNVL3.5-30B-A3B2026.042,056
NAEEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.112,054
MC-MoEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.112,037
DiEPExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.112,021
CAI-MoE (Token Drop)Backbone=KIMI-VL-A3B-INSTRUCT2026.042,013
Janus-Pro2026.051,978
MC-MoEExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.111,972
NAEEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.111,968
Top-k RoutingExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.111,964
VanillaBase Model=LLaVA-1.5-7B, Retained Tokens=576, Reduction Rate=0%2026.021,862
DiEPExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.111,838
Top-k RoutingExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.111,821
ApETBase Model=LLaVA-1.5-7B, Retained Tokens=192, Reduction Rate=66.7%2026.021,808
NAEEExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.111,802
ApETBase Model=LLaVA-1.5-7B, Retained Tokens=128, Reduction Rate=77.8%2026.021,801
PDropBase Model=LLaVA-1.5-7B, Retained Tokens=192, Reduction Rate=66.7%2026.021,797
VisionZipBase Model=LLaVA-1.5-7B, Retained Tokens=192, Reduction Rate=66.7%2026.021,783
VisionZipBase Model=LLaVA-1.5-7B, Retained Tokens=128, Reduction Rate=77.8%2026.021,763
PDropBase Model=LLaVA-1.5-7B, Retained Tokens=128, Reduction Rate=77.8%2026.021,761
SparseVLMBase Model=LLaVA-1.5-7B, Retained Tokens=192, Reduction Rate=66.7%2026.021,721
ApETBase Model=LLaVA-1.5-7B, Retained Tokens=64, Reduction Rate=88.9%2026.021,714
SparseVLMBase Model=LLaVA-1.5-7B, Retained Tokens=128, Reduction Rate=77.8%2026.021,696
VisionZipBase Model=LLaVA-1.5-7B, Retained Tokens=64, Reduction Rate=88.9%2026.021,690
Top-k RoutingExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.111,683
FastVBase Model=LLaVA-1.5-7B, Retained Tokens=192, Reduction Rate=66.7%2026.021,612
ToMeBase Model=LLaVA-1.5-7B, Retained Tokens=192, Reduction Rate=66.7%2026.021,563
PDropBase Model=LLaVA-1.5-7B, Retained Tokens=64, Reduction Rate=88.9%2026.021,561
SparseVLMBase Model=LLaVA-1.5-7B, Retained Tokens=64, Reduction Rate=88.9%2026.021,505
FastVBase Model=LLaVA-1.5-7B, Retained Tokens=128, Reduction Rate=77.8%2026.021,490
ToMeBase Model=LLaVA-1.5-7B, Retained Tokens=128, Reduction Rate=77.8%2026.021,343
FastVBase Model=LLaVA-1.5-7B, Retained Tokens=64, Reduction Rate=88.9%2026.021,256
ToMeBase Model=LLaVA-1.5-7B, Retained Tokens=64, Reduction Rate=88.9%2026.021,138