Long-Horizon Composite Skills on OSExpert-Eval
33GIMP Success RateOSExpert (Explore w/ Qwen-3-VL-8B, wo/ Fine Skills)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OSExpert (Explore w/ Qwen-3-VL-8B, wo/ Fine Skills)Agent Type=Agentic Framework2026.03 | 33 | 29 | |
| OSExpert (Explore w/ Qwen-3-VL-8B)Agent Type=Agentic Framework2026.03 | 33 | 29 | |
| OSExpert (Explore w/ GPT-5 & UI-TARS)Agent Type=Agentic Framework2026.03 | 33 | 31 | |
| Agent-S3 w/ GPT-5Agent Type=Agentic Framework2026.03 | 6 | 8 | |
| OpenCUA-7BAgent Type=Specialized Model2026.03 | 0 | 0 | |
| Computer-Use-PreviewAgent Type=Specialized Model2026.03 | 0 | 4 | |
| Qwen-3-VL-8BAgent Type=General Model2026.03 | 0 | 0 | |
| CoAct-1Agent Type=Agentic Framework2026.03 | 0 | 8 |