List your company
ResearchFormerly WizWand

Math Reasoning on MathVista (pass@1, avg@8)

79.6Avg Pass@8ADHint
43.252.6562.171.55Dec 15, 2025Jan 10, 2026Feb 5, 2026Mar 3, 2026Mar 29, 2026Apr 24, 2026May 21, 2026
Updated 5mo ago

Evaluation results

MethodLinks
ADHintBase Model=Qwen3-VL-8B-Instruct2025.1279.674
GRPOBase Model=Qwen3-VL-8B-Instruct2025.1278.874.6
ADHintBase MLLM=Qwen2.5-VL-7B2025.1274.874.4
GRPOBase MLLM=Qwen2.5-VL-7B2025.1273.473.4
SFT -> RLBase MLLM=Qwen2.5-VL-7B2025.1271.164.7
SFTBase MLLM=Qwen2.5-VL-7B2025.1268.759
VA-OPDTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0566.4—
Off-policy KDTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0565.3—
GRPOTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0564.6—
PAPOTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0564.5—
HintGRPOBase Model=Qwen3-VL-8B-Instruct2025.1264.464.9
CoT-SFTTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0564.4—
ADHintBase MLLM=Qwen2.5-VL-3B2025.1264.364.8
HintGRPOBase MLLM=Qwen2.5-VL-7B2025.1264.365.9
GHPOBase MLLM=Qwen2.5-VL-7B2025.1263.966.1
BaseTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0563.9—
OPDTeacher=Qwen3-VL-8B, Student=Qwen3-VL-2B, Training Data=Geometry3K, Temperature=1.0, Checkpoint Selection=Best checkpoint2026.0563.7—
SFT -> RLBase MLLM=Qwen2.5-VL-3B2025.1263.353.6
GRPOBase MLLM=Qwen2.5-VL-3B2025.1263.163.9
SFTBase MLLM=Qwen2.5-VL-3B2025.1261.351.9
LUFFYBase MLLM=Qwen2.5-VL-7B2025.1261.263.6
GHPOBase MLLM=Qwen2.5-VL-3B2025.1255.257.3
HintGRPOBase MLLM=Qwen2.5-VL-3B2025.1252.556.7
StepHintBase MLLM=Qwen2.5-VL-7B2025.1249.547.2
LUFFYBase MLLM=Qwen2.5-VL-3B2025.1245.641.2
StepHintBase MLLM=Qwen2.5-VL-3B2025.1244.634.9
Auxiliary Reward2026.03—71.2
Ctrl-RPower scaling factor (β)=0.22026.03—70.9
GRPO2026.03—68.1
OpenVLThinker2026.03—71.2
Trajectory-Hint2026.03—69.8