ResearchBenchmarksIndirect Object Identification on IOI evaluation episodes (held-out)Follow2.976Policy ScoreMechRL2.82722.90162.9763.0504May 25, 2026Evaluation ResultsMethodMethodLinksPolicy ScoreOracle ScoreGapMechRLK=1, In-episode picks=...K=1, In-episode picks=50, Evaluation episodes=≥ 202026.052.976—0.028OracleSelection=per-episode,...Selection=per-episode, Evaluation episodes=≥ 202026.05—2.948—