List your company
ResearchFormerly WizWand

Multimodal Understanding on MMMU (dev)

67.3AccuracyPVM-8B (SFT + GRPO)
18.10830.87943.6556.421Jan 24, 2026Feb 9, 2026Feb 25, 2026Mar 13, 2026Mar 29, 2026Apr 14, 2026May 1, 2026
Updated 6mo ago

Evaluation results

MethodLinks
PVM-8B (SFT + GRPO)Backbone=8B, Training Strategy=SFT + GRPO, Model Architecture=Persistent Visual Memory2026.0567.3
PVM-8B (SFT)Backbone=8B, Training Strategy=SFT, Model Architecture=Persistent Visual Memory2026.0566.7
PEARL-8BBackbone=8B, Training Strategy=RL-tuned2026.0565.3
Qwen3-VL-8B (LoRA-SFT + GRPO)Backbone=8B, Training Strategy=LoRA-SFT + GRPO2026.0564.7
ICoTTraining Strategy=Visual Injection2026.0563.3
Qwen3-VL-8B (LoRA-SFT)Backbone=8B, Training Strategy=LoRA-SFT2026.0563.3
PVM-4B (SFT + GRPO)Backbone=4B, Training Strategy=SFT + GRPO, Model Architecture=Persistent Visual Memory2026.0562.7
CoMemoTraining Strategy=Visual Injection2026.0562
Euclid-8BBackbone=8B, Training Strategy=RL-tuned2026.0562
OneThinker-8BBackbone=8B, Training Strategy=RL-tuned2026.0562
Qwen3-VL-8B (SFT)Backbone=8B, Training Strategy=SFT2026.0560.7
Qwen3-VL-8B (SFT + GRPO)Backbone=8B, Training Strategy=SFT + GRPO2026.0560.7
PVM-4B (SFT)Backbone=4B, Training Strategy=SFT, Model Architecture=Persistent Visual Memory2026.0560.7
MemVRTraining Strategy=Visual Injection2026.0559.3
Qwen3-VL-4B (SFT)Backbone=4B, Training Strategy=SFT2026.0558
Qwen3-VL-4B (SFT + GRPO)Backbone=4B, Training Strategy=SFT + GRPO2026.0558
Qwen3-VL-8B-InstructBackbone=8B2026.0557.3
Qwen3-VL-4B-InstructBackbone=4B2026.0557.3
Qwen3-VL-4B (LoRA-SFT)Backbone=4B, Training Strategy=LoRA-SFT2026.0556
Qwen3-VL-4B (LoRA-SFT + GRPO)Backbone=4B, Training Strategy=LoRA-SFT + GRPO2026.0556
DefenderIteration=32026.0125.33
DefenderIteration=22026.0123.33
Base (M_def^(0)) + Clean DataData=Cleaned2026.0121.33
Base (M_def^(0))2026.0120.67
DefenderIteration=12026.0120