GUI Navigation on OSWorld
28.2AccuracyBEAP-Agent
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BEAP-AgentModel=GPT-4o + UI-TARS-1.5-7B, Steps=502026.01 | 28.2 | 35.8 | 65.5 | 2.72 | |
| Agent S2Model=GPT-4o + UI-TARS-72B-DPO, Steps=502026.01 | 26.6 | — | — | — | |
| BEAP-Agent - w/o BacktrackModel=GPT-4o + UI-TARS-1.5-7B, Steps=502026.01 | 26.3 | — | — | — | |
| JEDIModel=GPT-4o + JEDI-7B, Steps=502026.01 | 25 | — | — | — | |
| UI-TARSModel=UI-TARS-1.5-7B, Steps=502026.01 | 24 | — | — | — | |
| BEAP-Agent - w/o TrackerModel=GPT-4o + UI-TARS-1.5-7B, Steps=502026.01 | 23.6 | — | — | — | |
| AGUVISModel=GPT-4o + AGUVIS-72B, Steps=-2026.01 | 17 | — | — | — | |
| Qwen2.5Model=Qwen2.5-vl-72B, Steps=-2026.01 | 8.8 | — | — | — | |
| OpenAIModel=GPT-4o, Steps=-2026.01 | 5 | — | — | — |