Conversational Starter Generation on Large-scale conversational system (Online A/B Test)
1.84Active Rate ChangeICEBREAKER
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ICEBREAKERBackbone=Doubao1.5-Lite, Optimization Strategy=Self-Reinforced Iterative Optimization2026.04 | 1.84 | 1.59 | 94.25 | 1.27 | |
| SFT + DPOBackbone=Doubao1.5-Lite, Training Type=Fine-tuned2026.04 | 1.16 | 0.42 | 56.41 | 0.68 | |
| SFTBackbone=Doubao1.5-Lite, Training Type=Fine-tuned2026.04 | 0.2 | 0.33 | 6.97 | -0.05 | |
| PEBackbone=Doubao1.5-Lite, Training Type=Training-free2026.04 | -0.01 | -0.26 | -16.16 | -0.17 |