Skill Generation on SkillsBench
22Baseline Success RateAutoRefine
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| AutoRefineTeacher Model=claude-sonnet-4.6, Teacher Trajectories=SPARK teacher trajectories2026.05 | 22 | 26.1 | 40.7 | 43.4 | 18.6 | 17.2 | 19 | 9 | |
| SPARKTeacher Model=claude-sonnet-4.6, Teacher Trajectories=SPARK teacher trajectories2026.05 | 22 | 24 | 50.8 | 53.1 | 28.8 | 29.1 | 21 | 3 | |
| EvoSkillTeacher Model=claude-sonnet-4.6, Teacher Trajectories=SPARK teacher trajectories2026.05 | 20.3 | 24.6 | 30.5 | 35.3 | 10.2 | 10.7 | 11 | 2 | |
| Trace2SkillTeacher Model=claude-sonnet-4.6, Teacher Trajectories=SPARK teacher trajectories2026.05 | 16.9 | 21.8 | 27.1 | 34.8 | 10.2 | 13 | 12 | 4 |