Task-Oriented Dialogue on Real-world Dialogue Benchmark (combined)
72.5Life Services TCBaseline
Evaluation results
| Method | Links | |||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BaselineTraining Steps=1000, Reward Supervision=Noisy2026.02 | 72.5 | 95 | 92.7 | 86.7 | 66.7 | 96.5 | 90.3 | 84.5 | 77.2 | 97.8 | 95.7 | 90.2 | 75.3 | 92 | 94.1 | 87.1 | 65.7 | 91.4 | 91 | 82.7 | 86.3 | |
| Reinforce++Training Steps=1000, Reward Supervision=Noisy2026.02 | 70.9 | 92.5 | 92.1 | 85.2 | 54.1 | 87.9 | 86 | 76 | 45.8 | 94.3 | 86.8 | 75.6 | 66.1 | 99.2 | 90.9 | 85.4 | 69.6 | 92.4 | 88.6 | 83.5 | 81.1 | |
| DFPOTraining Steps=1000, Reward Supervision=Noisy2026.02 | 70.8 | 94.4 | 91.8 | 85.7 | 82.1 | 94.5 | 94.9 | 90.5 | 75.4 | 98.8 | 94.9 | 89.7 | 71.1 | 96.8 | 91.3 | 86.4 | 69.4 | 92.6 | 90.9 | 84.3 | 87.3 | |
| GRPOTraining Steps=1000, Reward Supervision=Noisy2026.02 | 47.4 | 14.6 | 25.7 | 29.2 | 41 | 20.5 | 28.5 | 30 | 24.6 | 21.6 | 28.5 | 24.9 | 43.3 | 10 | 23.5 | 25.6 | 50.9 | 12.7 | 24.5 | 29.4 | 27.8 | |
| PPOTraining Steps=1000, Reward Supervision=Noisy2026.02 | 43.8 | 63.4 | 70.4 | 59.2 | 46.2 | 77.1 | 70.8 | 64.7 | 31.6 | 77.4 | 79 | 62.7 | 47.4 | 61.5 | 70.2 | 59.7 | 45.3 | 70.4 | 68.7 | 61.5 | 61.5 | |
| Dr.GRPOTraining Steps=1000, Reward Supervision=Noisy2026.02 | 34.5 | 45.3 | 54.2 | 44.7 | 37 | 53 | 60 | 49.9 | 23 | 66 | 62 | 50.2 | 44 | 50 | 58 | 50.8 | 35 | 54 | 51 | 46.6 | 48.4 |