List your company
ResearchFormerly WizWand

Multimodal Reasoning on MathVista (Accuracy, AvgLen, Ratio)

85.9AccuracyQwen3-VL-32B-Thinking
33.79647.32360.8574.377Sep 30, 2025Nov 13, 2025Dec 28, 2025Feb 11, 2026Mar 27, 2026May 11, 2026Jun 25, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Qwen3-VL-32B-ThinkingParameters=32B, Thinking Mode=true2026.0185.9——
Qwen3-VL-8B-ThinkingParameters=8B, Thinking Mode=true2026.0181.4——
EvoCUA-32BParameters=32B2026.0180.4——
Token-ReplayModel=Qwen3-VL, #Params=32B2026.0679.4——
PRCRModel=Qwen3-VL, #Params=32B2026.0679.2——
BaselineModel=Qwen3-VL, #Params=32B2026.0678.5——
RLSDBase Model=Qwen3-VL-8B-Instruct, Max Context Length=81922026.0478.1——
GRPOBase Model=Qwen3-VL-8B-Instruct, Max Context Length=81922026.0476.2——
GRPO + OPSDBase Model=Qwen3-VL-8B-Instruct, Max Context Length=81922026.0475.9——
EvoCUA-8BParameters=8B2026.0175.8——
OPSDBase Model=Qwen3-VL-8B-Instruct, Max Context Length=81922026.0475.1——
TGRL-DAPOData=39K, Base Model=Qwen-2.5-VL-7B2026.0375——
PRCRModel=InternVL3.5, #Params=8B2026.0674.6——
NoisyRolloutData=6.4K2026.0374.5——
Token-ReplayModel=InternVL3.5, #Params=8B2026.0674.3——
Perception-R1Data=1.4K2026.0374.2——
SDPOBase Model=Qwen3-VL-8B-Instruct, Max Context Length=81922026.0474——
Base LLMBase Model=Qwen3-VL-8B-Instruct, Max Context Length=81922026.0473.8——
VL-RethinkerTool=✗, Param Size=7B2025.1173.7——
BaselineModel=InternVL3.5, #Params=8B2026.0673.7——
TGRL-GRPOData=39K, Base Model=Qwen-2.5-VL-7B2026.0373.2——
PRCRModel=InternVL3.5, #Params=14B2026.0673——
PRCRModel=Qwen3-VL, #Params=8B2026.0672.8——
ThinkLiteData=11K2026.0372.7——
Token-ReplayModel=InternVL3.5, #Params=14B2026.0672.7——
MM-EurekaTool=✗, Param Size=7B2025.1172.6——
Token-ReplayModel=Qwen3-VL, #Params=8B2026.0672.6——
ThinkLite-VL-7BTraining Protocol=Zero RL2025.1271.9——
DeepEyesV2Tool=General, Param Size=7B2025.1171.9——
BaselineModel=InternVL3.5, #Params=14B2026.0671.9——
VLAA-ThinkerTool=✗, Param Size=7B2025.1171.7——
InternVL3Tool=✗, Param Size=8B2025.1171.6——
ThinkLiteTool=✗, Param Size=7B2025.1171.6——
OpenVLThinkerData=50K2026.0371.5——
TGRL-GRPOData=2.1K, Base Model=Qwen-2.5-VL-7B2026.0371.5——
BaselineModel=Qwen3-VL, #Params=8B2026.0671.5——
TGRL-DAPOData=2.1K, Base Model=Qwen-2.5-VL-7B2026.0371.4——
DAPOData=39K, Base Model=Qwen-2.5-VL-7B2026.0371.3——
MM-Eureka-Qwen-7BTraining Protocol=Zero RL2025.1271.2——
MSSRModel=Qwen2.5-VL-7B, Training Protocol=Zero RL2025.1271.1——
OpenCUA-72BParameters=72B2026.0170.9——
GRPOData=39K, Base Model=Qwen-2.5-VL-7B2026.0370.9——
GRPOData=2.1K, Base Model=Qwen-2.5-VL-7B2026.0370.5——
GPT-4.1VLM Category=Proprietary VLMs2025.0970.4——
InternVL3-8BVLM Category=Open-source VLMs2025.0970.4——
DAPOData=2.1K, Base Model=Qwen-2.5-VL-7B2026.0370.3——
DeepEyesTool=Crop, Param Size=7B2025.1170.1——
DeepEyes-7BVLM Category=Tool-Calling VLMs2025.0970.1——
OpenVLThinker-7BTraining Protocol=SFT + RL2025.1270——
GRPOModel=Qwen2.5-VL-7B, Training Protocol=Zero RL2025.1270——
ThymeTool=Code, Param Size=7B2025.1170——
Thyme-VL-7BVLM Category=Tool-Calling VLMs2025.0970——
EvoCUA-OpenCUA-72BBase Model=OpenCUA-72B2026.0169.4——
RLOOModel=Qwen2.5-VL-7B, Training Protocol=Zero RL2025.1269.2——
DeepSketcher-7BVLM Category=Inner Visual Thought VLMs2025.0969.1——
VLAA-ThinkerData=25K2026.0368.8——
REINFORCE++Model=Qwen2.5-VL-7B, Training Protocol=Zero RL2025.1268.5——
Visual-PRMPRM Trainable Params=7B2026.0668.5——
Qwen-2.5-VLTool=✗, Param Size=7B2025.1168.3——
Qwen2.5-VL-7BVLM Category=Open-source VLMs2025.0968.2——
VLAA-Thinker-7BTraining Protocol=SFT + RL2025.1268——
Qwen-2.5-VL-7BData=-2026.0367.5——
MMS-PRMPRM Trainable Params=0M2026.0667.5——
Qwen2.5-VL-7BTraining Protocol=Zero RL2025.1267.2——
Claude3.7-SonnetVLM Category=Proprietary VLMs2025.0966.8——
VL-PRM300KPRM Trainable Params=8B2026.0665.7——
Mulberry-Qwen-7BVLM Category=Open-source VLMs2025.0965.1——
URSAPRM Trainable Params=8B2026.0664.9——
Bagel-Zebra-CoT-7BVLM Category=Inner Visual Thought VLMs2025.0964.7——
R1-OnevisionData=165K2026.0363.9——
Mirage-7BVLM Category=Inner Visual Thought VLMs2025.0963.7——
MSSRModel=Qwen2.5-VL-3B, Training Protocol=Zero RL2025.1263——
R1-Onevision-7BTraining Protocol=SFT + RL2025.1262.9——
GRPOModel=Qwen2.5-VL-3B, Training Protocol=Zero RL2025.1261.7——
SVIPPRM Trainable Params=7B2026.0661.6——
RLOOModel=Qwen2.5-VL-3B, Training Protocol=Zero RL2025.1259.7——
Qwen2.5-VL-3BTraining Protocol=Zero RL2025.1259.5——
LLaVA-OVTool=✗, Param Size=7B2025.1158.6——
VILASR-7BVLM Category=Tool-Calling VLMs2025.0957.6——
MathCoder-VLVLM Category=Open-source VLMs2025.0955.1——
REINFORCE++Model=Qwen2.5-VL-3B, Training Protocol=Zero RL2025.1247.7——
Prune-on-LogicBackbone=Qwen3-VL, Evaluation Protocol=SFT2026.0245.9299.56.5
No CompressionBackbone=Qwen3-VL, Evaluation Protocol=SFT2026.0245.7511.611.2
XMCCBackbone=Qwen3-VL, Evaluation Protocol=SFT2026.0245.3912
StepEntropyBackbone=Qwen3-VL, Evaluation Protocol=SFT2026.0243.9251.35.7
XMCCBackbone=Qwen2.5-VL, Evaluation Protocol=SFT2026.0237.689.72.4
Prune-on-LogicBackbone=Qwen2.5-VL, Evaluation Protocol=SFT2026.0237.3324.28.7
StepEntropyBackbone=Qwen2.5-VL, Evaluation Protocol=SFT2026.0236.8232.26.3
No CompressionBackbone=Qwen2.5-VL, Evaluation Protocol=SFT2026.0235.8524.114.6