List your company
ResearchFormerly WizWand

Multimodal mathematical reasoning on MathVista (test)

74.2AccuracyQwen-VL-Max
57.5661.8866.270.52Jan 7, 2026Jan 20, 2026Feb 2, 2026Feb 16, 2026Mar 1, 2026Mar 14, 2026Mar 28, 2026
Updated 7mo ago

Evaluation results

MethodLinks
Qwen-VL-MaxModel Category=Closed-Source Models2026.0174.2—
GPRO-7BBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0174.2115.3
FAST-7BBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0173.8120.7
Vision-R1Base Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0173.5—
GSPO + KAWHI (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0372.155.54
GRPO + KAWHI (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0371.851.98
DAPO + KAWHI (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0371.2153.06
GRPO (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=GRPO2026.0371.0751.17
MM-R1Base Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0171185.6
GSPO (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=GSPO2026.0370.554.14
Qwen3-VL-4B-Instruct BaseBackbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=none, RL Method=Base2026.0370.4350.53
OpenVLThinkerBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0170.2305.7
GSPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=GSPO2026.0369.4251.4
DAPO (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=DAPO2026.0369.451.8
GSPO + KAWHIBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0369.3752.94
GRPO + KAWHIBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0369.351
DAPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=DAPO2026.0369.250.79
DAPO + KAWHIBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0369.0851.85
VPPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=fine-grained reward, RL Method=VPPO2026.0368.4250.29
GRPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=GRPO2026.0368.3750.18
Qwen2.5-VL-7BBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0168.2189.1
Claude-3.5 SonnetModel Category=Closed-Source Models2026.0167.7—
FT-RLBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=fine-grained reward, RL Method=FT-RL2026.0367.6849.37
Qwen2.5-VL-7B-Instruct BaseBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=none, RL Method=Base2026.0367.4747.68
Step-GRPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=fine-grained reward, RL Method=Step-GRPO2026.0366.8747.82
GPRO-3BBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0166.8145.2
FAST-3BBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0166.2158.7
R1-OneVisionBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0164.1402.5
GPT-4oModel Category=Closed-Source Models2026.0163.8—
LMM-R1Base Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0163.2245
Qwen2.5-VL-3BBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0162.3212.9
MulberryBase Model Scale=7B, Base Architecture=Qwen2-VL2026.0162.1275
Curr-ReFTBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0161.995.9
Qwen2-VL-7BBase Model Scale=7B, Base Architecture=Qwen2-VL2026.0158.2265.9