E2W-Bench Counting and Location Reasoning 1.0
61E2W-1 ScoreCoRL
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CoRLBackbone=Qwen2.5VL-7B, Protocol=+CoRL (Ours)2026.03 | 61 | 97 | 90 | 82.67 | |
| CoRLBackbone=Qwen2.5VL-3B, Protocol=+CoRL (Ours)2026.03 | 59 | 75.5 | 86 | 73.5 | |
| Qwen2.5VL-3B (+SFT)Backbone=Qwen2.5VL-3B, Protocol=+SFT2026.03 | 47 | 63 | 84 | 64.67 | |
| Qwen2.5VL-7B (+SFT)Backbone=Qwen2.5VL-7B, Protocol=+SFT2026.03 | 44.5 | 88 | 84.5 | 72.33 | |
| GPT-5Model Category=Closed Source Models2026.03 | 42.5 | 48.5 | 72.5 | 54.5 | |
| Doubao-Seed-1.6Model Category=Closed Source Models2026.03 | 35 | 40 | 46 | 40.33 | |
| GLM-4.5vModel Category=Open Source Models2026.03 | 34.5 | 29 | 56 | 39.83 | |
| Gemini-2.5-ProModel Category=Closed Source Models2026.03 | 32.5 | 42.5 | 50 | 41.67 | |
| Qwen2.5VL-3B (+RL-ZERO)Backbone=Qwen2.5VL-3B, Protocol=+RL-ZERO2026.03 | 23 | 39.5 | 83.5 | 48.67 | |
| Qwen2.5VL-3B (Zero-shot)Backbone=Qwen2.5VL-3B, Protocol=Zero-shot2026.03 | 22 | 15.5 | 58 | 31.83 | |
| SpaceQwen-3BModel Category=Open Source Models2026.03 | 21.5 | 6 | 60 | 29.17 | |
| Qwen2.5VL-32BModel Category=Open Source Models2026.03 | 21.5 | 28 | 37 | 28.83 | |
| Qwen2.5VL-7B (Zero-shot)Backbone=Qwen2.5VL-7B, Protocol=Zero-shot2026.03 | 17 | 17 | 64.5 | 32.83 | |
| LLaMA-3.2-11bModel Category=Open Source Models2026.03 | 16.5 | 12.5 | 17.5 | 15.5 | |
| Qwen2.5VL-7B (+RL-ZERO)Backbone=Qwen2.5VL-7B, Protocol=+RL-ZERO2026.03 | 16 | 56 | 82.5 | 51.5 |