List your company
ResearchFormerly WizWand

Visual Reasoning on MMVP

86.3AccuracyGPT-4o
22.23638.86855.572.132Jun 13, 2024Oct 16, 2024Feb 18, 2025Jun 23, 2025Oct 26, 2025Feb 28, 2026Jul 3, 2026
Updated 3mo ago

Evaluation results

MethodLinks
GPT-4oVisual Sketchpad=Enabled2024.0686.3
GPT-4oVisual Sketchpad=Disabled2024.0685.3
One-step Latent Pred.Reasoning Paradigm=Latent-Space Reasoning, Reasoning Space=Latent2026.0779.33
ProLaViT (Full)Reasoning Paradigm=Progressive Latent Visual Thought, Reasoning Space=Latent, Training Strategy=Distance-Weighted Diversity Loss2026.0779
CoVTReasoning Paradigm=Latent-Space Reasoning, Reasoning Space=Latent2026.0778.33
ProLaViT + L_div^marginReasoning Paradigm=Progressive Latent Visual Thought, Reasoning Space=Latent2026.0778.33
ProLaViT (Base)Reasoning Paradigm=Progressive Latent Visual Thought, Reasoning Space=Latent2026.0778
CoT SFTReasoning Paradigm=Text-Space Reasoning, Reasoning Space=Text2026.0777.66
Qwen2.5-VL-InstructReasoning Paradigm=Base Model, Reasoning Space=N/A2026.0777.33
LVRReasoning Paradigm=Latent-Space Reasoning, Reasoning Space=Latent2026.0777.3
SFTReasoning Paradigm=Text-Space Reasoning, Reasoning Space=Text2026.0777
LLaVA-OneVisionReasoning Paradigm=Zero-Shot VLMs2026.0174
RIS+VLPOBackbone=Qwen2.5-VL-7B, Latent Tokens=5, Optimization=Visual-latent Policy Optimization (VLPO)2026.0573.76
RISBackbone=Qwen2.5-VL-7B, Latent Tokens=52026.0573.55
GPT-4 TurboVisual Sketchpad=Enabled2024.0673.3
Unsilencing Latent ReasoningBackbone=Qwen2.5VL-7B2026.0573
Vision-R1Reasoning Paradigm=Tool-use & RL Enhanced Reasoning2026.0172.67
VL-RethinkerReasoning Paradigm=Tool-use & RL Enhanced Reasoning2026.0172.67
CoVTBackbone=Qwen2.5VL-7B2026.0572.67
LaserReasoning Paradigm=Latent Reasoning2026.0172
GPT-4oModel Type=Proprietary2026.0572
LVRBackbone=Qwen2.5-VL-7B, Latent Tokens=52026.0571.03
GPT-4 TurboVisual Sketchpad=Disabled2024.0671
DeepEyesReasoning Paradigm=Tool-use & RL Enhanced Reasoning2026.0170
DMLRBackbone=Qwen2.5VL-7B2026.0570
MonetBackbone=Qwen2.5-VL-7B, Latent Tokens=52026.0570
LVRRFBackbone=Qwen2.5VL-7B2026.0569.67
ICoTBackbone=Qwen2.5VL-7B2026.0569.33
CCoTBackbone=Qwen2.5VL-7B2026.0569
LVRBackbone=Qwen2.5VL-7B2026.0569
GPT-4oReasoning Paradigm=Zero-Shot VLMs2026.0168.7
PAPOReasoning Paradigm=Tool-use & RL Enhanced Reasoning2026.0168.67
VanillaBackbone=Qwen2.5VL-7B2026.0568.67
MonetReasoning Paradigm=Latent Reasoning2026.0168
MCoTBackbone=Qwen2.5VL-7B2026.0568
Qwen2.5-VL-7B+GLSDBackbone=Qwen2.5-VL-7B, Training=Grounded Latent Supervision Dataset (GLSD)2026.0566.8
Qwen2.5-VL-7BReasoning Paradigm=Zero-Shot VLMs2026.0165.67
MonetBackbone=Qwen2.5VL-7B2026.0565.67
LVRReasoning Paradigm=Latent Reasoning2026.0164
Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.0563.49
Unsilencing Latent ReasoningBackbone=Qwen2.5VL-3B2026.0559.33
CoVTBackbone=Qwen2.5-VL-7B, Latent Tokens=52026.0558.7
LVRBackbone=Qwen2.5VL-3B2026.0558.67
LVRRFBackbone=Qwen2.5VL-3B2026.0558.33
InternVL3.5-8BReasoning Paradigm=Zero-Shot VLMs2026.0157.67
DMLRBackbone=Qwen2.5VL-3B2026.0556.67
ICoTBackbone=Qwen2.5VL-3B2026.0556
VanillaBackbone=Qwen2.5VL-3B2026.0555.67
CCoTBackbone=Qwen2.5VL-3B2026.0555.33
MCoTBackbone=Qwen2.5VL-3B2026.0554.33
SOTA [10]2024.0649.3
LEADBase Model=OpenVLThinker-7B, Decoding Strategy=LEAD2026.0347.2
LEADBase Model=VL-Rethinker-7B, Decoding Strategy=LEAD2026.0346.6
OpenVLThinker-7BDecoding Strategy=Standard2026.0346.5
LEADBase Model=Vision-R1-7B, Decoding Strategy=LEAD2026.0346.3
LEADBase Model=R1-Onevision-7B, Decoding Strategy=LEAD2026.0345
MemVRBase Model=R1-Onevision-7B, Decoding Strategy=MemVR2026.0344.5
Vision-R1-7BDecoding Strategy=Standard2026.0344
SIDBase Model=R1-Onevision-7B, Decoding Strategy=SID2026.0343.2
R1-Onevision-7BDecoding Strategy=Standard2026.0343
VCDBase Model=R1-Onevision-7B, Decoding Strategy=VCD2026.0342.9
VL-Rethinker-7BDecoding Strategy=Standard2026.0342
LEADBase Model=VL-Cogito-7B, Decoding Strategy=LEAD2026.0342
Gemini-Pro2024.0640.7
VL-Cogito-7BDecoding Strategy=Standard2026.0340
GPT-4V-preview2024.0638.7
LLaVA-1.5-13B2024.0624.7