List your company
ResearchFormerly WizWand

Mathematical Reasoning on MathVision (Accuracy)

83.9AccuracyQwen3.5
12.722431.201249.6868.1588May 21, 2025Jul 28, 2025Oct 4, 2025Dec 12, 2025Feb 18, 2026Apr 27, 2026Jul 5, 2026
Updated 3mo ago

Evaluation results

MethodLinks
Qwen3.5Parameters=35BA3B, Internal Reasoning (Think mode)=false2026.0583.9
Gemini3-proModel Category=Closed-Source SOTA Models2026.0583.4
SenseNova-U1Parameters=30BA3B, Internal Reasoning (Think mode)=true2026.0579.63
Qwen3.5Parameters=9B, Internal Reasoning (Think mode)=false2026.0578.9
SenseNova-U1Parameters=8B, Internal Reasoning (Think mode)=true2026.0575.82
Qwen3-VL-235B-ThinkingModel Category=Open-Source Large Baselines2026.0574.6
GPT-5-20250807Model Category=Closed-Source SOTA Models2026.0572
Gemma4Parameters=26BA4B, Internal Reasoning (Think mode)=false2026.0568.95
Doubao-Seed-1.6Model Category=Closed-Source SOTA Models2026.0567.8
Qwen3VLParameters=30BA3B, Internal Reasoning (Think mode)=true2026.0565.7
GLM-4.5VModel Category=Closed-Source SOTA Models2026.0565.6
LongCat-NextParameters=68BA3B, Internal Reasoning (Think mode)=false2026.0564.7
Qwen3VLParameters=8B, Internal Reasoning (Think mode)=true2026.0562.7
OSTBase Model=Qwen3-VL-8B-Instruct, Sampling Ratio=Best-20%2026.0558.5
DEITABase Model=Qwen3-VL-8B-Instruct, Sampling Ratio=Top-20%2026.0558.3
Qwen3-VL-8B-Instruct + LLM-as-a-JudgeSampling Ratio=20%2026.0558.2
Qwen3-VL-8B-InstructStrategy=Base2026.0557.8
Qwen3-VL-8B-Instruct + RandomSampling Ratio=20%2026.0557.1
Kimi-vl-A3B-thinkingModel Category=Open-Source Large Baselines2026.0556.8
OSTBase Model=Qwen3-VL-4B-Instruct, Sampling Ratio=Best-20%2026.0554.8
Qwen3-VL-8B-Instruct + Full SFTSampling Ratio=100%2026.0554.8
Qwen3-VL-4B-Instruct + LLM-as-a-JudgeSampling Ratio=20%2026.0554.6
DEITABase Model=Qwen3-VL-4B-Instruct, Sampling Ratio=Top-20%2026.0554.5
Qwen3-VL-4B-InstructStrategy=Base2026.0554.1
Qwen3-VL-4B-Instruct + RandomSampling Ratio=20%2026.0554
Qwen3-VL-4B-Instruct + Full SFTSampling Ratio=100%2026.0553.8
Qwen3-VL-8BFine-tuning setting=CPO2026.0751.61
Gemini-2.0-FlashActivation Replay=false2025.1147.8
Qwen3-VL-4B-Instruct / GRPOStudent=Qwen3-VL-4B-Instruct, Teacher=Self, Method=GRPO2026.0746.7
Qwen3-VL-4B-Instruct / H-OPD (8B)Student=Qwen3-VL-4B-Instruct, Teacher=Qwen3-VL-8B + Qwen3-8B, Method=H-OPD2026.0746.7
Qwen3-VL-4B-Instruct / ExOPD (8B)Student=Qwen3-VL-4B-Instruct, Teacher=Qwen3-VL-8B, Method=ExOPD2026.0746.4
Qwen3-VL-4B-Instruct / OPD (8B)Student=Qwen3-VL-4B-Instruct, Teacher=Qwen3-VL-8B, Method=OPD2026.0746.1
Qwen3-VL-8BType=Baseline, Evaluation Tool=VLMEvalKit2026.0745.4
Qwen3-VL-4BType=Baseline, Evaluation Tool=VLMEvalKit2026.0744.7
Qwen3-VL-4B-Instruct / SFT (8B)Student=Qwen3-VL-4B-Instruct, Teacher=Qwen3-VL-8B, Method=SFT2026.0742.1
Claude-3.5 SonnetActivation Replay=false2025.1141.3
Qwen3-VL-2B-Instruct / H-OPD (8B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-8B + Qwen3-8B, Method=H-OPD2026.0736.8
Qwen3-VL-8BFine-tuning setting=Base2026.0736.78
MM-Eureka-Qwen-32BActivation Replay=true2025.1135.5
MM-Eureka-Qwen-32BActivation Replay=false2025.1135.2
Qwen3-VL-2B-Instruct + LLM-as-a-JudgeSampling Ratio=20%2026.0535
QvQ-72B-PreviewActivation Replay=false2025.1134.9
DEITABase Model=Qwen3-VL-2B-Instruct, Sampling Ratio=Top-20%2026.0534.9
OSTBase Model=Qwen3-VL-2B-Instruct, Sampling Ratio=Best-20%2026.0534.8
TemplateRLBackbone=Qwen2.5-VL-3B-Instruct, Batch size=128, Samples per question=16, Template guidance quantity=22025.0534.6
Qwen3-VL-2B-InstructStrategy=Base2026.0534.5
Qwen3-VL-2B-Instruct + RandomSampling Ratio=20%2026.0534.3
Qwen3-VL-2B-Instruct + Full SFTSampling Ratio=100%2026.0534.1
MMR1-Math-v0-7BActivation Replay=true2025.1133.9
Qwen3-VL-2B-Instruct / H-OPD (4B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-4B + Qwen3-4B, Method=H-OPD2026.0733.5
VL-Rethinker-7BActivation Replay=true2025.1133.2
MMR1-Math-v0-7BActivation Replay=false2025.1132.6
Qwen3-VL-8BFine-tuning setting=GRPO2026.0731.78
Qwen3-VL-2B-Instruct / ExOPD (4B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-4B, Method=ExOPD2026.0731.6
MM-Eureka-Qwen-7BActivation Replay=true2025.1131.5
GPT-4oActivation Replay=false2025.1131.1
GRPOBackbone=Qwen2.5-VL-3B-Instruct, Batch size=128, Samples per question=162025.0530.9
Qwen3-VL-2B-Instruct / OPD (4B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-4B, Method=OPD2026.0730.9
MM-Eureka-Qwen-7BActivation Replay=false2025.1130.6
Qwen3-VL-2B-Instruct / ExOPD (8B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-8B, Method=ExOPD2026.0730.6
VL-Rethinker-7BActivation Replay=false2025.1130.3
Qwen3-VL-2B-Instruct / GRPOStudent=Qwen3-VL-2B-Instruct, Teacher=Self, Method=GRPO2026.0730.3
Qwen3-VL-2B-Instruct / OPD (8B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-8B, Method=OPD2026.0730.3
InternVL3-8BActivation Replay=false2025.1128.6
Qwen3-VL-2BType=Baseline, Evaluation Tool=VLMEvalKit2026.0728.6
Qwen3-VL-2B-Instruct / SFT (8B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-8B, Method=SFT2026.0728.2
CoTBackbone=Qwen2.5-VL-3B-Instruct, Batch size=128, Samples per question=162025.0527.7
GPT-4o-miniActivation Replay=false2025.1127.3
Qwen3-VL-2B-Instruct / SFT (4B)Student=Qwen3-VL-2B-Instruct, Teacher=Qwen3-VL-4B, Method=SFT2026.0726.8
Qwen3-VL-8BFine-tuning setting=GSPO2026.0726.35
+BETAPRMSelector=+BETAPRM, Backbone=InternVL3-14B, Best-of-N strategy=Best-of-162026.0525.66
+BETAPRMSelector=+BETAPRM, Backbone=InternVL2.5-8B, Best-of-N strategy=Best-of-162026.0525.66
Qwen2.5-VL-7BActivation Replay=false2025.1125.5
+BETAPRMSelector=+BETAPRM, Backbone=InternVL3-8B, Best-of-N strategy=Best-of-162026.0524.34
+BETAPRMSelector=+BETAPRM, Backbone=Qwen2.5-VL-7B, Best-of-N strategy=Best-of-162026.0524.34
Qwen3-VL-8BFine-tuning setting=LoRA2026.0723.98
+Standard PRMSelector=+Standard PRM, Backbone=InternVL3-14B, Best-of-N strategy=Best-of-162026.0523.03
+Standard PRMSelector=+Standard PRM, Backbone=InternVL3-8B, Best-of-N strategy=Best-of-162026.0522.69
Kimi-VL-16BActivation Replay=false2025.1121.8
+Standard PRMSelector=+Standard PRM, Backbone=InternVL2.5-8B, Best-of-N strategy=Best-of-162026.0521.38
+Standard PRMSelector=+Standard PRM, Backbone=Qwen2.5-VL-7B, Best-of-N strategy=Best-of-162026.0521.38
+Base (w/o training)Selector=+Base (w/o training), Backbone=InternVL2.5-8B, Best-of-N strategy=Best-of-162026.0520.72
+Base (w/o training)Selector=+Base (w/o training), Backbone=InternVL3-14B, Best-of-N strategy=Best-of-162026.0519.74
Qwen3-VL-8BFine-tuning setting=FFT2026.0718.82
+Base (w/o training)Selector=+Base (w/o training), Backbone=InternVL3-8B, Best-of-N strategy=Best-of-162026.0518.75
Single PassSelector=Single Pass, Backbone=InternVL2.5-8B, Best-of-N strategy=1 (Single Pass)2026.0518.08
LLaVA-OV-7BActivation Replay=false2025.1117
Qwen2.5-VL-3BActivation Replay=false2025.1116.1
+Base (w/o training)Selector=+Base (w/o training), Backbone=Qwen2.5-VL-7B, Best-of-N strategy=Best-of-162026.0515.46