Web Navigation on WebVoyager OOD
56.3AllrecipesVanilla
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| VanillaModel=GPT-4o2026.06 | 56.3 | 53.7 | 56.6 | 60.5 | 57.7 | 65.1 | 56.9 | 65.2 | 59 | |
| VanillaModel=Claude 3 Opus2026.06 | 45.9 | 58.6 | 58.1 | 55 | 56.9 | 68.2 | 55.3 | 51.5 | 56.1 | |
| PEEU-SFTModel=Qwen2.5-VL-7B, Training Trajectories=2k2026.06 | 17.7 | 53.6 | 16.2 | 25.2 | 19.5 | 35.7 | 48.7 | 28.2 | 30.6 | |
| Atomic-SFTModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 15.5 | 17 | 11.6 | 23.2 | 0 | 7.1 | 4.8 | 19.5 | 12.3 | |
| Atomic-SFTModel=Qwen2.5-VL-3B, Training Trajectories=2k2026.06 | 13.3 | 26.8 | 4.6 | 23.2 | 7.3 | 9.5 | 17 | 32.6 | 16.7 | |
| Atomic-SFTModel=Qwen2.5-VL-7B, Training Trajectories=2k2026.06 | 13.3 | 51.2 | 6.9 | 25.5 | 0 | 9.5 | 39 | 28.2 | 21.7 | |
| PEEU-SFTModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 8.8 | 24.3 | 18.6 | 16.2 | 7.3 | 16.6 | 7.3 | 26 | 15.6 | |
| PEEU-SFTModel=Qwen2.5-VL-3B, Training Trajectories=2k2026.06 | 8.8 | 46.3 | 13.9 | 13.9 | 9.7 | 14.2 | 21.9 | 30.4 | 19.8 | |
| VanillaModel=Qwen2.5-VL-72B2026.06 | 6.6 | 58.5 | 25.5 | 32.5 | 17 | 21.4 | 36.5 | 36.9 | 29.3 | |
| PEEU-GRPOModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 6.6 | 24.3 | 3 | 23.2 | 9.7 | 7.1 | 0 | 15.2 | 11.1 | |
| Coarse-SFTModel=Qwen2.5-VL-7B, Training Trajectories=2k2026.06 | 6.6 | 34.1 | 20.9 | 20.9 | 17 | 14.2 | 17 | 21.7 | 19 | |
| Trajectory-PromptModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 4.4 | 0 | 0 | 4.6 | 2.4 | 9.5 | 2.4 | 6.5 | 3.7 | |
| PEEU-GRPOModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 4.4 | 26.8 | 18.6 | 20.9 | 21.9 | 33.3 | 12.1 | 21.7 | 19.9 | |
| Atomic-SFTModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 2.2 | 2.4 | 0 | 4.6 | 7.3 | 7.1 | 0 | 15.2 | 4.8 | |
| PEEU-SFTModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 2.2 | 7.3 | 6.9 | 11.6 | 2.4 | 0 | 4.8 | 10.8 | 5.7 | |
| VanillaModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 2.2 | 7.3 | 9.3 | 4.6 | 9.7 | 16.6 | 0 | 13 | 7.8 | |
| Atomic-PromptModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 2.2 | 0 | 6.9 | 4.6 | 2.4 | 9.5 | 0 | 4.3 | 3.7 | |
| Atomic-GRPOModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 2.2 | 19.5 | 0 | 18.6 | 0 | 11.9 | 0 | 28.2 | 10 | |
| VanillaModel=Qwen2.5-VL-7B, Training Trajectories=2k2026.06 | 2.2 | 7.3 | 9.3 | 4.6 | 9.7 | 16.6 | 0 | 13 | 7.8 | |
| VanillaModel=Qwen2.5-VL-32B2026.06 | 0 | 39 | 16.2 | 32.5 | 21.9 | 19 | 34.1 | 19.5 | 22.7 | |
| VanillaModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2.1 | 0.2 | |
| Atomic-PromptModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |
| Trajectory-PromptModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |
| Coarse-SFTModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 0 | 0 | 0 | 2.3 | 2.4 | 4.7 | 2.4 | 4.3 | 2 | |
| Coarse-GRPOModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 0 | 2.4 | 0 | 20.9 | 0 | 2.3 | 2.4 | 17.3 | 5.6 | |
| Atomic-GRPOModel=Qwen2.5-VL-3B, Training Trajectories=0.1k2026.06 | 0 | 12.1 | 2.3 | 11.6 | 0 | 9.5 | 0 | 8.6 | 5.5 | |
| Coarse-SFTModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 0 | 4.8 | 0 | 4.6 | 0 | 7.1 | 4.8 | 17.3 | 4.8 | |
| Coarse-GRPOModel=Qwen2.5-VL-7B, Training Trajectories=0.1k2026.06 | 0 | 17 | 7.1 | 20.9 | 4.8 | 4.7 | 12.1 | 26 | 11.5 | |
| VanillaModel=Qwen2.5-VL-3B, Training Trajectories=2k2026.06 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2.1 | 0.2 | |
| Coarse-SFTModel=Qwen2.5-VL-3B, Training Trajectories=2k2026.06 | 0 | 12.1 | 6.9 | 6.9 | 9.7 | 14.2 | 17 | 39.1 | 13.2 |