ResearchBenchmarksDialog on Real-world Dialog Domains AggregateFollow87.31Average AccuracyDFPO25.440441.502757.56573.6273Feb 5, 2026Evaluation ResultsMethodMethodLinksAverage AccuracyDFPOTraining Steps=1000Training Steps=10002026.0287.31BaselineTraining Steps=1000Training Steps=10002026.0286.26Reinforce++Training Steps=1000Training Steps=10002026.0281.15PPOTraining Steps=1000Training Steps=10002026.0261.55Dr.GRPOTraining Steps=1000Training Steps=10002026.0248.42GRPOTraining Steps=1000Training Steps=10002026.0227.82