Windows UI Navigation on WindowsAgentArena (WAA)
74.5Success RateHuman Performance
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Human Performance2026.01 | 74.5 | — | |
| CUA-Skill AgentBackbone=GPT-5, Decoding Strategy=Best of 32026.01 | 57.5 | 30 | |
| AgentS3Backbone=GPT-5, Decoding Strategy=Best-of-32026.01 | 56.6 | 100 | |
| CUA-Skill AgentBackbone=GPT-52026.01 | 50.3 | 30 | |
| AgentS3Backbone=GPT-52026.01 | 49 | 50 | |
| RoTS-32BMax Steps=502026.05 | 39.1 | — | |
| Operator2026.01 | 37.4 | 50 | |
| RoTS-32BMax Steps=152026.05 | 35.9 | — | |
| Jedi-7B w/ GPT4-oMax Steps=502026.05 | 32.9 | — | |
| UFO-2Backbone=GPT-o12026.01 | 30.5 | 30 | |
| Jedi-7B w/ GPT4-oMax Steps=152026.05 | 30.2 | — | |
| Agent-S2Backbone=Claude-3.5-Sonnet2026.01 | 29.8 | 15 | |
| RoTS-7BMax Steps=502026.05 | 28.2 | — | |
| RoTS-7BMax Steps=152026.05 | 24.9 | — | |
| ScaleCUA-32BMax Steps=502026.05 | 24.2 | — | |
| ScaleCUA-32BMax Steps=152026.05 | 21.4 | — | |
| Ultra-CUA-7BBackbone=QWen2.5VL-FT2026.01 | 21.1 | 15 | |
| ScaleCUA-7BMax Steps=502026.05 | 20.7 | — | |
| STEVEBackbone=GPT4o2026.01 | 20.1 | 40 | |
| NAVIBackbone=GPT4o2026.01 | 19.5 | 15 | |
| Agent-SBackbone=GPT-4o2026.01 | 18.2 | 15 | |
| UITARS-1.5-7BBackbone=Qwen2.5VL-FT2026.01 | 18.1 | 15 | |
| ScaleCUA-7BMax Steps=152026.05 | 18 | — | |
| UI-TARS-72B-DPOMax Steps=502026.05 | 17.9 | — | |
| UI-TARS-1.5-7BMax Steps=502026.05 | 15.9 | — | |
| UI-TARS1Backbone=Qwen2VL-FT2026.01 | 15.7 | 50 | |
| OpenCUA-7BMax Steps=152026.05 | 13.5 | — | |
| Qwen2.5-VL-72BMax Steps=152026.05 | 11.8 | — | |
| UI-TARS-1.5-7BMax Steps=152026.05 | 11.1 | — | |
| UI-TARS-72B-DPOMax Steps=152026.05 | 11.1 | — | |
| Qwen2.5-VL-72BMax Steps=502026.05 | 9.7 | — | |
| Claude 3.7 SonnetMax Steps=152026.05 | 7.1 | — | |
| Claude 3.7 SonnetMax Steps=502026.05 | 6.4 | — |