Plan-quality verification on τ-bench retail 20 held-out tasks
17.2CheckBeforeSubmit (CBS)Offline AW
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Offline AWEvaluation Tasks=16 training / 20 held-out tasks, Protocol=Harness MDP scoring protocol, Rubric=Adapted plan-quality rubric2026.07 | 17.2 | 18.2 |