List your company
ResearchFormerly WizWand

Multi-discipline Multimodal Understanding on CMMMU

25.3Overall AccuracyMAGIC
20.41221.68122.9524.219May 25, 2026
Updated 5mo ago

Evaluation results

MethodLinks
MAGICTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0525.3——
ICONSTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0525.2——
PerplexityTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0523.9——
SemDeDupTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0522.8——
Full-FinetuneTraining Budget=100%, Backbone=LLaVA-1.5-13B2026.0522.1——
RandomTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0521.9——
D2-PruningTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0521.9——
EL2NTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0521.7——
COINCIDETraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0521.4——
RDSTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0521.3——
Self-SupTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0521——
CLIP-ScoreTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0520.8——
Self-FilterTraining Budget=20%, Backbone=LLaVA-1.5-13B2026.0520.6——
CLIPLLM=Qwen2-7B, Vision Tower=ViT-L/142024.07—33.772.03
CLIPLLM=Qwen2-72B, Vision Tower=ViT-L/142024.07—37.4576.63
MLCDLLM=Qwen2-7B, Vision Tower=ViT-L/142024.07—33.173.88
MLCDLLM=Qwen2-72B, Vision Tower=ViT-L/142024.07—41.278.59