Diagram Understanding on AI2D lite
82.8AccuracyPVM-8B (SFT + GRPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| PVM-8B (SFT + GRPO)Backbone=8B, Training Strategy=SFT + GRPO, Model Architecture=Persistent Visual Memory2026.05 | 82.8 | |
| Euclid-8BBackbone=8B, Training Strategy=RL-tuned2026.05 | 82.6 | |
| PEARL-8BBackbone=8B, Training Strategy=RL-tuned2026.05 | 81.8 | |
| OneThinker-8BBackbone=8B, Training Strategy=RL-tuned2026.05 | 81.4 | |
| Qwen3-VL-8B (LoRA-SFT + GRPO)Backbone=8B, Training Strategy=LoRA-SFT + GRPO2026.05 | 81 | |
| PVM-4B (SFT + GRPO)Backbone=4B, Training Strategy=SFT + GRPO, Model Architecture=Persistent Visual Memory2026.05 | 81 | |
| PVM-8B (SFT)Backbone=8B, Training Strategy=SFT, Model Architecture=Persistent Visual Memory2026.05 | 80.8 | |
| PVM-4B (SFT)Backbone=4B, Training Strategy=SFT, Model Architecture=Persistent Visual Memory2026.05 | 80 | |
| Qwen3-VL-8B-InstructBackbone=8B2026.05 | 79.8 | |
| Qwen3-VL-8B (LoRA-SFT)Backbone=8B, Training Strategy=LoRA-SFT2026.05 | 79.8 | |
| CoMemoTraining Strategy=Visual Injection2026.05 | 79.6 | |
| Qwen3-VL-8B (SFT + GRPO)Backbone=8B, Training Strategy=SFT + GRPO2026.05 | 79.6 | |
| Qwen3-VL-4B (LoRA-SFT)Backbone=4B, Training Strategy=LoRA-SFT2026.05 | 79.2 | |
| Qwen3-VL-8B (SFT)Backbone=8B, Training Strategy=SFT2026.05 | 79 | |
| MemVRTraining Strategy=Visual Injection2026.05 | 78.8 | |
| ICoTTraining Strategy=Visual Injection2026.05 | 78.6 | |
| Qwen3-VL-4B (SFT + GRPO)Backbone=4B, Training Strategy=SFT + GRPO2026.05 | 78.6 | |
| Qwen3-VL-4B-InstructBackbone=4B2026.05 | 78.4 | |
| Qwen3-VL-4B (SFT)Backbone=4B, Training Strategy=SFT2026.05 | 77.6 | |
| Qwen3-VL-4B (LoRA-SFT + GRPO)Backbone=4B, Training Strategy=LoRA-SFT + GRPO2026.05 | 76.8 |