Multimodal GUI-agent action execution on Android in the Wild (AITW) held-out 500-example (test)
6.6AITW AccuracyWEASEL greedy selection
Evaluation Results
| Method | Links | |
|---|---|---|
| WEASEL greedy selectionSubset size=3.1K, Base policy model=Qwen2.5-VL-3B-Instruct, Diversity embedding=SigLIP2-Large-384, Importance model=CLIP ViT-L/14-3362026.05 | 6.6 | |
| Random selectionSubset size=3.1K, Base policy model=Qwen2.5-VL-3B-Instruct2026.05 | 5.8 | |
| Qwen2.5-VL-3B-Instructtype=base model2026.05 | 4.4 |