List your company
ResearchFormerly WizWand

Multimodal Reasoning on MMMU-Pro (Acc, # Tokens)

58.4AccuracyLaRe
18.994429.224739.45549.6853Oct 24, 2025Dec 3, 2025Jan 12, 2026Feb 21, 2026Apr 2, 2026May 12, 2026Jun 22, 2026
Updated 4mo ago

Evaluation results

MethodLinks
LaReBase VLM=Qwen3-VL-4B-Instruct2025.1158.4103.5
CCoTBase VLM=Qwen3-VL-4B-Instruct2025.1155.5234.9
LVR*Base VLM=Qwen3-VL-4B-Instruct2025.1153.9144.4
CoF*Base VLM=Qwen3-VL-4B-Instruct2025.1153.7269.1
Standard CoTBase VLM=Qwen3-VL-4B-Instruct2025.1152.4226.8
ICoTBase VLM=Qwen3-VL-4B-Instruct2025.1152.3211.7
MCOUTBase VLM=Qwen3-VL-4B-Instruct2025.1152124.5
MM-CoTBase VLM=Qwen3-VL-4B-Instruct2025.1149.2204.2
LaReBase VLM=Qwen3-VL-2B-Instruct2025.1144.2102.4
STRIDEBase model=Qwen2.5-VL-7B2026.0644.2—
SRPOBase model=Qwen2.5-VL-7B2026.0642.3—
CoF*Base VLM=Qwen3-VL-2B-Instruct2025.1141.2274
ICoTBase VLM=Qwen3-VL-2B-Instruct2025.1140.8214.7
MM-CoTBase VLM=Qwen3-VL-2B-Instruct2025.1139.8198.8
MCOUTBase VLM=Qwen3-VL-2B-Instruct2025.1139120.1
LVR*Base VLM=Qwen3-VL-2B-Instruct2025.1138.9134.7
MM-EurekaBase model=Qwen2.5-VL-7B2026.0637.6—
Standard CoTBase VLM=Qwen3-VL-2B-Instruct2025.1137224.2
CCoTBase VLM=Qwen3-VL-2B-Instruct2025.1135.5203.5
RECAPBackbone=Qwen2.5-VL-7B, Training=finetuned2025.1034.15—
Reasoning-onlyBackbone=Qwen2.5-VL-7B, Training=finetuned2025.1033.87—
UniformBackbone=Qwen2.5-VL-7B, Training=finetuned2025.1031.91—
CoresetBackbone=Qwen2.5-VL-7B, Training=finetuned2025.1031.91—
PropMixBackbone=Qwen2.5-VL-7B, Training=finetuned2025.1031.39—
BaseBase model=Qwen2.5-VL-7B2026.0630.5—
LwFBackbone=Qwen2.5-VL-7B, Training=finetuned2025.1029.59—
CFPOGBackbone=Qwen3-VL-2B-Thinking2026.0627.96—
Vision-R1-7B2025.1026.76—
VLAA-Thinker-7B2025.1026.3—
Qwen2.5-VL-7Bvariant=Base model2025.1025.55—
PAPOGBackbone=Qwen3-VL-2B-Thinking2026.0623.42—
OpenVLThinker-7B2025.1021.79—
GRPOBackbone=Qwen3-VL-2B-Thinking2026.0620.51—