List your company
ResearchFormerly WizWand

Multimodal mathematical reasoning on MathVision (test)

60.3AccuracyOpenAI-o1
14.95626.72838.550.272Jun 8, 2025Jul 26, 2025Sep 13, 2025Nov 1, 2025Dec 20, 2025Feb 7, 2026Mar 28, 2026
Updated 7mo ago

Evaluation results

MethodLinks
OpenAI-o1#Data=/2025.0660.3——
Claude-3.7-Sonnet#Data=/2025.0641.3——
Qwen-VL-MaxModel Category=Closed-Source Models2026.0139.3——
Claude-3.5 SonnetModel Category=Closed-Source Models2026.0137.9——
Qwen2.5-VL-72B-IT#Data=/2025.0637.9——
GSPO (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=GSPO2026.0332.89—54.14
GSPO + KAWHIBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0332.57—52.94
DAPO + KAWHIBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0331.91—51.85
GPRO-7BBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0131.2195.6—
GSPO + KAWHI (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0331.09—55.54
FAST-7BBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0130.6204.8—
GPT-4oModel Category=Closed-Source Models2026.0130.4——
GPT-4o#Data=/2025.0630.4——
MM-R1Base Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0130.2324.6—
R1-OneVisionBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0129.9692.8—
OpenVLThinkerBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0129.6457.2—
GRPO + KAWHIBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0329.28—51
Perception-R1-7B#Data=1.4K2025.0628.6——
DAPO + KAWHI (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0328.43—53.06
VPPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=fine-grained reward, RL Method=VPPO2026.0328.42—50.29
GRPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=GRPO2026.0328.29—50.18
DAPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=DAPO2026.0328.21—50.79
DAPO (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=DAPO2026.0327.63—51.8
MM-Eureka-7B#Data=15K2025.0627.6——
GSPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=uniform reward, RL Method=GSPO2026.0327.3—51.4
GPRO-3BBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0127.1298.6—
FAST-3BBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0126.8323.5—
VLAA-Thinker-7B#Data=25K2025.0626.7——
SophiaVL-R1-7B#Data=130K2025.0626.6——
FT-RLBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=fine-grained reward, RL Method=FT-RL2026.0326.12—49.37
Qwen2.5-VL-7BBase Model Scale=7B, Base Architecture=Qwen2.5-VL2026.0125.6443—
LMM-R1Base Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0125.2447.8—
Qwen2.5-VL-7B-IT#Data=/2025.0625.1——
Step-GRPOBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=fine-grained reward, RL Method=Step-GRPO2026.0324.76—47.82
OpenVLThinker-7B#Data=25K2025.0624.3——
VirgoBase Model Scale=7B, Base Architecture=Qwen2-VL2026.0124——
MulberryBase Model Scale=7B, Base Architecture=Qwen2-VL2026.0123.4349.2—
Qwen2.5-VL-7B-Instruct BaseBackbone Model=Qwen2.5-VL-7B-Instruct, Reward Strategy=none, RL Method=Base2026.0323.36—47.68
GRPO + KAWHI (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=KAWHI2026.0323.03—51.98
GRPO (Qwen3)Backbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=uniform reward, RL Method=GRPO2026.0322.16—51.17
R1-OneVision-7B#Data=155K2025.0621.9——
Qwen3-VL-4B-Instruct BaseBackbone Model=Qwen3-VL-4B-Instruct, Reward Strategy=none, RL Method=Base2026.0321.38—50.53
Qwen2.5-VL-3BBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0121.2450.6—
Curr-ReFTBase Model Scale=3B, Base Architecture=Qwen2.5-VL2026.0120.1240.1—
InternVL2.5-8B#Data=/2025.0619.7——
Qwen2-VL-7BBase Model Scale=7B, Base Architecture=Qwen2-VL2026.0118.8443—
Qwen2-VL-7B-IT#Data=/2025.0616.7——