ResearchBenchmarksInduction on Induction evaluation episodes (held-out)Follow3.408Policy ScoreMechRL3.23763.32283.4083.4932May 25, 2026Evaluation ResultsMethodMethodLinksPolicy ScoreOracle ScorePerformance GapMechRLK=1, In-episode picks=...K=1, In-episode picks=50, Evaluation episodes=≥ 202026.053.408—0.006OracleSelection=per-episode,...Selection=per-episode, Evaluation episodes=≥ 202026.05—3.402—