ResearchBenchmarksMulti-turn tool-use on τ³-airline 20-task (held-out)Follow60.2Pass@1SGCD30.5638.25545.9553.645Jun 10, 2026Evaluation ResultsMethodMethodLinksPass@1Action Success RateInformation Success RateSGCDstep=120step=1202026.0660.243.593.8GRPO+KLstep=120step=1202026.0658.34191.7SFT basestep=800step=8002026.0643.826.968.8SDPOstep=300step=3002026.0631.7091.7