End-to-end task success on OSWorld 361 tasks
52.1Success RateS3 + IntentScore
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| S3 + IntentScoreNumber of candidates (N)=3, Horizon length=50, Reasoning backbone=GPT-5-mini, Coding-based actions=disabled2026.04 | 52.1 | 46.6 | |
| S3 + Alignment onlyNumber of candidates (N)=3, Horizon length=50, Reasoning backbone=GPT-5-mini, Coding-based actions=disabled2026.04 | 48 | 40.1 | |
| S3 + LLM self-evaluationNumber of candidates (N)=3, Horizon length=50, Reasoning backbone=GPT-5-mini, Coding-based actions=disabled2026.04 | 45.8 | 52.1 | |
| S3 w/ N=3(baseline)Number of candidates (N)=3, Horizon length=50, Reasoning backbone=GPT-5-mini, Coding-based actions=disabled2026.04 | 45.2 | — | |
| S3 + random selectionNumber of candidates (N)=3, Horizon length=50, Reasoning backbone=GPT-5-mini, Coding-based actions=disabled2026.04 | 44.2 | 56.9 |