Visual Equivalence Reward Modeling on Visual-ERM-Bench AVG
42.1F1h ScoreVisual-ERM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Visual-ERMCategory=Open-source, Backbone=Qwen3-VL-8B-Instruct2026.03 | 42.1 | 44.7 | 58.4 | |
| Gemini-3-FlashCategory=Proprietary2026.03 | 40.6 | 43.4 | 53.4 | |
| Gemini-2.5-ProCategory=Proprietary2026.03 | 37.8 | 40.9 | 59.1 | |
| GPT-5.2Category=Proprietary2026.03 | 32.7 | 35 | 58.9 | |
| Qwen3-VL-235B-InstructCategory=Open-source2026.03 | 29.5 | 32.4 | 56.2 | |
| GPT-4oCategory=Proprietary2026.03 | 25 | 29.5 | 56.5 | |
| InternVL3.5-8BCategory=Open-source2026.03 | 6.7 | 9.6 | 32.5 | |
| Qwen3-VL-8B-InstructCategory=Open-source2026.03 | 5.3 | 6.5 | 17.5 | |
| Qwen2.5-VL-7BCategory=Open-source2026.03 | 2.8 | 5.1 | 15.2 |