List your company
ResearchFormerly WizWand

Real-world Visual Understanding on RealWorldQA

81.4AccuracyInternVL2.5-8B-GenRecal
56.54462.99769.4575.903Jun 18, 2025Aug 11, 2025Oct 4, 2025Nov 27, 2025Jan 20, 2026Mar 15, 2026May 8, 2026
Updated 5mo ago

Evaluation results

MethodLinks
InternVL2.5-8B-GenRecalTeacher VLM=InternVL2.5-78B2025.0681.4
Qwen2.5-VL-72B + CARESCost=-82%2025.1079
InternVL2.5-8B-GenRecalTeacher VLM=Qwen2-VL-72B2025.0678.8
InternVL2.5-78B2025.0678.7
Qwen2-VL-72B2025.0677.8
Qwen2.5-VL-72B2025.1077
SkillGraphModel=Qwen3-VL 8B-Instruct, Baseline=Complete2026.0476.4
SkillGraphModel=Qwen3-VL 8B-Instruct, Baseline=Random2026.0476.1
InternVL2.5-8B-GenRecalTeacher VLM=InternVL2-76B2025.0675.5
GPT-4o (0513)2025.0675.4
SkillGraphModel=Qwen3-VL 8B-Instruct, Baseline=Layered2026.0475.4
RandomModel=Qwen3-VL 8B-Instruct2026.0475.3
CompleteModel=Qwen3-VL 8B-Instruct2026.0475.2
SkillGraphModel=Qwen3-VL 8B-Instruct, Baseline=Centralized2026.0474.9
SkillGraphModel=InternVL3-8B, Baseline=Complete2026.0474.7
SkillGraphModel=Qwen3-VL 8B-Instruct, Baseline=Linear2026.0474.4
SkillGraphModel=InternVL3-8B, Baseline=Random2026.0474.1
LayeredModel=Qwen3-VL 8B-Instruct2026.0473.9
Vanilla MoEBackbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0473.72
Molmo-72B2025.0673.7
SkillGraphModel=InternVL3-8B, Baseline=Layered2026.0473.7
MACS (Ours)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0473.58
LinearModel=Qwen3-VL 8B-Instruct2026.0473.4
SkillGraphModel=InternVL3-8B, Baseline=Linear2026.0473.4
SkillGraphModel=InternVL3-8B, Baseline=Centralized2026.0473.4
CentralizedModel=Qwen3-VL 8B-Instruct2026.0473.2
MACS (w/o Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0473.15
RandomModel=InternVL3-8B2026.0473.1
CompleteModel=InternVL3-8B2026.0472.9
InternVL2.5-8B-GenRecalTeacher VLM=NVLM-72B2025.0672.5
InternVL2-76B2025.0672.2
SkillGraphModel=Qwen2.5-VL 7B-Instruct, Baseline=Random2026.0472.2
LayeredModel=InternVL3-8B2026.0472.2
LLaVA-OneVision-72B2025.0671.9
SkillGraphModel=Qwen2.5-VL 7B-Instruct, Baseline=Complete2026.0471.9
LinearModel=InternVL3-8B2026.0471.6
CentralizedModel=InternVL3-8B2026.0471.5
DirectAnswerModel=Qwen3-VL 8B-Instruct2026.0471.4
Qwen3-VL-SegModel Size=4B, Training Stage=S-22026.0571.2
SkillGraphModel=Qwen2.5-VL 7B-Instruct, Baseline=Layered2026.0471.1
SkillGraphModel=Qwen2.5-VL 7B-Instruct, Baseline=Centralized2026.0470.9
Qwen3-VLModel Size=4B, Training Stage=instruct2026.0570.9
DirectAnswerModel=InternVL3-8B2026.0470.7
SkillGraphModel=Qwen2.5-VL 7B-Instruct, Baseline=Linear2026.0470.6
SkillGraphModel=LLaVA-OV Qwen2-7B, Baseline=Complete2026.0470.5
RandomModel=Qwen2.5-VL 7B-Instruct2026.0470.4
SkillGraphModel=LLaVA-OV Qwen2-7B, Baseline=Random2026.0470.3
CompleteModel=Qwen2.5-VL 7B-Instruct2026.0470.3
SkillGraphModel=LLaVA-OV Qwen2-7B, Baseline=Layered2026.0470.1
NVLM-72B2025.0669.9
SkillGraphModel=LLaVA-OV Qwen2-7B, Baseline=Centralized2026.0469.8
LayeredModel=Qwen2.5-VL 7B-Instruct2026.0469.7
SkillGraphModel=LLaVA-OV Qwen2-7B, Baseline=Linear2026.0469.6
CAI-MoE (Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0469.46
RandomModel=LLaVA-OV Qwen2-7B2026.0469.4
LinearModel=Qwen2.5-VL 7B-Instruct2026.0469.3
CompleteModel=LLaVA-OV Qwen2-7B2026.0469.2
CentralizedModel=Qwen2.5-VL 7B-Instruct2026.0469.2
DirectAnswerModel=Qwen2.5-VL 7B-Instruct2026.0468.6
LayeredModel=LLaVA-OV Qwen2-7B2026.0468.5
LinearModel=LLaVA-OV Qwen2-7B2026.0468.1
Vanilla MoEBackbone=KIMI-VL-A3B-INSTRUCT2026.0468.07
InternVL3-8B2025.1068
InternVL3-8B + CARESCost=-82%2025.1068
InternVL3-8B + CARES-ARCost=-82%2025.1068
Qwen3-VLModel Size=4B, Training Stage=S-12026.0568
MACS (Ours)Backbone=KIMI-VL-A3B-INSTRUCT2026.0467.96
CentralizedModel=LLaVA-OV Qwen2-7B2026.0467.9
CAI-MoE (Token Drop)Backbone=QWEN3-VL-30B-A3B-INSTRUCT2026.0467.89
MACS (w/o Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.0467.63
Gemini-1.5-Pro2025.0667.5
DirectAnswerModel=LLaVA-OV Qwen2-7B2026.0467.2
InternVL-3.5Model Size=4B2026.0566.3
InternVL2-8BModel=InternVL2-8B, Pruning Method=None2025.0965.5
PTPModel=InternVL2-8B, Pruning Method=PTP, Pruning Ratio (r)=0.5, Balancing Weight (alpha)=0.52025.0965.3
CAI-MoE (Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.0465.18
Vanilla MoEBackbone=INTERNVL3.5-30B-A3B2026.0464.87
Top-k RoutingExpert skipping ratio (ρ)=0, Number of experts (k)=8, Backbone=InternVL-3.5-30B-A3B-HF2025.1164.77
MACS (Ours)Backbone=INTERNVL3.5-30B-A3B2026.0464.69
MACS (w/o Expanded)Backbone=INTERNVL3.5-30B-A3B2026.0464.12
CAI-MoE (Token Drop)Backbone=KIMI-VL-A3B-INSTRUCT2026.0463.42
MoDESExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1163.1
DiEPExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1162.56
Gemma3-27BModel Size=27B2025.1262.5
MoDESExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1162.48
MC-MoEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1162.13
MoDESExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.1161.73
NAEEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1161.47
Gemma3-4B + AuditDMModel Size=4B, Auditing Component=AuditDM2025.1261.4
Top-k RoutingExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.1161.34
CAI-MoE (Expanded)Backbone=INTERNVL3.5-30B-A3B2026.0461.05
GPT-4o2025.1061
GPT-4o + CARESCost=-84%2025.1061
MC-MoEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1160.28
Claude-3.5-Sonnet2025.0660.1
NAEEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1159.47
DiEPExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.1158.74
CAI-MoE (Token Drop)Backbone=INTERNVL3.5-30B-A3B2026.0458.34
Gemma3-12BModel Size=12B2025.1258.3
PTPModel=InternVL2-2B, Pruning Method=PTP, Pruning Ratio (r)=0.5, Balancing Weight (alpha)=0.52025.0957.5