Visual Probing on VisualProbe Hard
44.3AccuracyPERIA-8B
Evaluation Results
| Method | Links | |
|---|---|---|
| PERIA-8BModel Type=Proposed2026.06 | 44.3 | |
| Mini-o3Model Type=Tool-Augmented2026.06 | 41.5 | |
| Gemini-2.5-FlashModel Type=Commercial2026.06 | 33.6 | |
| Gemini-2.5-ProModel Type=Commercial2026.06 | 32 | |
| w/o RLModel Type=Ablation2026.06 | 32 | |
| Qwen3-235BModel Type=Open-Source2026.06 | 27.3 | |
| Qwen3-8BModel Type=Backbone Baseline2026.06 | 26.6 | |
| w/o SFTModel Type=Ablation2026.06 | 26.4 | |
| GPT-5Model Type=Commercial2026.06 | 25.4 | |
| Qwen3-30BModel Type=Open-Source2026.06 | 25.4 | |
| Qwen3-32BModel Type=Open-Source2026.06 | 24.5 | |
| VTool-R1-7BModel Type=Tool-Augmented2026.06 | 22.6 | |
| InternVL3.5-14BModel Type=Open-Source2026.06 | 19.8 | |
| w/o ToolModel Type=Ablation2026.06 | 19.8 | |
| R1-OneVision-7BModel Type=Tool-Augmented2026.06 | 17.3 | |
| InternVL3.5-8BModel Type=Open-Source2026.06 | 15.2 |