Visual Equivalence Reward Modeling on Visual-ERM-Bench Table
56.4F1 Score (h)Visual-ERM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Visual-ERMCategory=Open-source, Backbone=Qwen3-VL-8B-Instruct2026.03 | 56.4 | 57.6 | 74.8 | |
| Gemini-3-FlashCategory=Proprietary2026.03 | 48.1 | 50.1 | 45.6 | |
| Gemini-2.5-ProCategory=Proprietary2026.03 | 46.4 | 48 | 49.9 | |
| GPT-5.2Category=Proprietary2026.03 | 39.3 | 40.6 | 54.6 | |
| Qwen3-VL-235B-InstructCategory=Open-source2026.03 | 35.7 | 37.4 | 56.2 | |
| GPT-4oCategory=Proprietary2026.03 | 32.9 | 35.7 | 49.5 | |
| InternVL3.5-8BCategory=Open-source2026.03 | 9.9 | 10.9 | 31.7 | |
| Qwen3-VL-8B-InstructCategory=Open-source2026.03 | 7 | 7.8 | 21.4 | |
| Qwen2.5-VL-7BCategory=Open-source2026.03 | 2.3 | 3.1 | 12.6 |