List your company
ResearchFormerly WizWand

Mathematical Reasoning on MathVista (Accuracy)

89.2AccuracyGemini 3-Pro
69.887274.901179.91584.9289Jun 18, 2025Aug 13, 2025Oct 8, 2025Dec 3, 2025Jan 28, 2026Mar 25, 2026May 21, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Gemini 3-Pro2026.0289.2——
Gemini3-proModel Category=Closed-Source SOTA Models2026.0588.5——
Qwen3-VLmode=Thinking2026.0286.8——
Doubao-Seed-1.6Model Category=Closed-Source SOTA Models2026.0585.9——
Qwen3-VL-235B-ThinkingModel Category=Open-Source Large Baselines2026.0585.9——
TwC (Ours) - Img & TxtCategory=Think with Comic, Note=G-t-R2026.0285.8——
ERNIE 5.02026.0284.8——
GLM-4.5VModel Category=Closed-Source SOTA Models2026.0584.6——
ERNIE 5.0-BaseModel type=pre-trained2026.0284.4——
Qwen3-VL-32B2025.0683.8——
Gemini 2.5-Pro2026.0282.7——
GPT-5tier=High2026.0282.1——
InternVL3.5-38B2025.0681.9——
GPT-5-Thinking*Model Category=Close-source, Reference Source=OpenCompass leaderboard2025.0981.9——
GPT-5-20250807Model Category=Closed-Source SOTA Models2026.0581.9——
GenRecal (InternVL3.5-8B)Teacher VLM=Qwen3-VL-32B2025.0681.5——
GenRecal (InternVL3.5-8B)Teacher VLM=InternVL3.5-38B2025.0681.3——
Gemini-2.5-Pro*Model Category=Close-source, Reference Source=OpenCompass leaderboard2025.0980.9——
GenRecal (Qwen3-VL-8B)Teacher VLM=InternVL3.5-38B2025.0680.3——
GenRecal (Qwen3-VL-8B)Teacher VLM=Qwen3-VL-32B2025.0680.1——
Kimi-vl-A3B-thinkingModel Category=Open-Source Large Baselines2026.0580.1——
InternVL3.5-8B2025.0678.4——
Qwen3-VL-8B-Instruct + LLM-as-a-JudgeSampling Ratio=20%2026.0577.6——
DEITABase Model=Qwen3-VL-8B-Instruct, Sampling Ratio=Top-20%2026.0577.5——
Penguin-VLModel Size=8B2026.0377.4——
OSTBase Model=Qwen3-VL-8B-Instruct, Sampling Ratio=Best-20%2026.0577.4——
Qwen3-VL-8B2025.0677.2——
Qwen3-VLModel Size=8B2026.0377.2——
Qwen3-VL-8B-InstructStrategy=Base2026.0577.2——
HEEDPipeline Stage=C4+, Post-training=SFT+DPO2026.0577.2——
VL-Cogito-7B + LEADBackbone=VL-Cogito-7B, Decoding Strategy=LEAD2026.0376.3——
Qwen3-VL-8B-Instruct + RandomSampling Ratio=20%2026.0576.1——
TeacherPipeline Stage=C0, Post-training=SFT+DPO2026.0576——
OSTBase Model=Qwen3-VL-4B-Instruct, Sampling Ratio=Best-20%2026.0575.8——
VL-Rethinker-7B + LEADBackbone=VL-Rethinker-7B, Decoding Strategy=LEAD2026.0375.6——
VAPO-Thinker-7BModel Category=Our models, Parameter Scale=7B2025.0975.6——
RSAPipeline Stage=C3+, Post-training=SFT+DPO2026.0575.4——
Qwen3-VL-4B-Instruct + LLM-as-a-JudgeSampling Ratio=20%2026.0575.3——
ThinkLite-VLData Size=11k2026.0475.1——
DEITABase Model=Qwen3-VL-4B-Instruct, Sampling Ratio=Top-20%2026.0575.1——
TwC (Ours) - Only ImageCategory=Think with Comic, Note=direct2026.0275——
InternVL2.5-8B-GenRecalTeacher VLM=InternVL2.5-78B2025.0674.9——
Vision-R1-7B + LEADBackbone=Vision-R1-7B, Decoding Strategy=LEAD2026.0374.9——
VL-Rethinker-7BBackbone=VL-Rethinker-7B2026.0374.9——
VL-RethinkerData Size=39k2026.0474.9——
V-STARData Size=40k2026.0474.9——
Qwen3-VL-4B-InstructStrategy=Base2026.0574.9——
VL-Cogito-7BBackbone=VL-Cogito-7B2026.0374.8——
VL-CogitoData Size=80k2026.0474.8——
Qwen2.5-VL-32BParameters=32B2026.0274.7——
GeoFocusModel Scale=7B2026.0274.3——
InternVL-3.5Model Size=8B2026.0374.2——
Qwen3-VL-4B-Instruct + RandomSampling Ratio=20%2026.0574.2——
RuCL2026.0274.1——
HSAPipeline Stage=C2+, Post-training=SFT+DPO2026.0574.1——
MM-Eureka-Qwen-32BActivation Replay=true2025.1174——
OpenAI-o1Training Paradigm=Closed-source2026.0573.9——
Vision-R1-7BCategory=Open-source Reasoning Models2026.0173.5——
Vision-R1-7BBackbone=Vision-R1-7B2026.0373.5——
Vision-R1-7BModel Category=Open-source, Parameter Scale=7B2025.0973.5——
Vision-R1Data Size=210k2026.0473.5——
MM-Eureka-Qwen-7BActivation Replay=true2025.1173.5——
Vision-R1-7BTraining Paradigm=RL-trained, Parameters=7B2026.0573.5——
Qwen2.5-VL-7B-Instruct + Faithful-MR1Backbone model=Qwen2.5-VL-7B-Instruct, Training Data=19.2K2026.0573.5——
KDPipeline Stage=C1+, Post-training=SFT+DPO2026.0573.4——
ThinkLite-VL-7BParameters=7B2026.0273.3——
GazeVLM (Ours)Gaze Bias=Enabled, Base Model=Qwen3-VL-4B2026.0573.3——
VL-Rethinker-7BParameters=7B2026.0273.27——
Qwen3-VL-4B-Instruct + Full SFTSampling Ratio=100%2026.0573.2——
Perception-R1-7BTraining Data=1.4K2026.0573.2——
MM-Eureka-7BParameters=7B2026.0273——
MM-Eureka-Qwen-7BActivation Replay=false2025.1173——
Perception-R1-7BParameters=7B2026.0272.8——
VL-RethinkerParadigm=Tool-use & RL Enhanced Reasoning, Data Size=39K2026.0472.8——
AutoToolSize=7B2026.0572.8——
Claude-Sonnet 4.5Category=MLLM, Note=direct2026.0272.5——
VL-Rethinker-7BActivation Replay=true2025.1172.4——
InternVL2.5-78B2025.0672.3——
OpenVLThinkerData Size=59.2k2026.0472.3——
GRPOModel Scale=7B2026.0272.1——
MM-Eureka-Qwen-32BActivation Replay=false2025.1172.1——
VL-Rethinker-7BActivation Replay=false2025.1172——
InternVL2.5-8B-GenRecalTeacher VLM=InternVL2-76B2025.0671.9——
DeepEyesV2Tool Use=true, Param Size=7B2026.0371.9——
MMR1-Math-v0-7BActivation Replay=true2025.1171.9——
InternVL2.5-38BParameters=38B2026.0271.84——
InternVL3.5-2BThinking Mode=true2025.1271.8——
Qwen3-VL-4BTraining Paradigm=Vanilla Open-source, Parameters=4B2026.0571.7——
InternVL3Tool Use=false, Param Size=8B2026.0371.6——
Semantic-backTool Use=false, Param Size=7B2026.0371.6——
InternVL3.5-8BParadigm=Zero-Shot VLMs, Data Size=70K2026.0471.6——
DeepEyesSize=7B2026.0571.6——
Gemini-3-ProCategory=MLLM, Note=direct2026.0271.5——
InternVL2.5-8B-GenRecalTeacher VLM=Qwen2-VL-72B2025.0671.4——
OpenVLThinker-7BParameters=7B2026.0271.38——
VRETool Use=false, Param Size=7B2026.0371.2——
Video-R1-7BCategory=Open-source Reasoning Models2026.0171——
MMR1-Math-v0-7BActivation Replay=false2025.1171——
Qwen2.5-VL-7B-Instruct + GRPOBackbone model=Qwen2.5-VL-7B-Instruct, Training Data=19.2K2026.0570.9——
Vision-R1-7BParameters=7B2026.0270.63——