Query Rewriting on Production session logs short-video search engine (Offline evaluation)
1.231Average RewardPhase 4
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Phase 4MPO Phase=4th semi-online phase2026.06 | 1.231 | 1.256 | 58.9 | |
| Phase 3MPO Phase=3rd semi-online phase2026.06 | 1.192 | 1.227 | 56.1 | |
| Phase 2MPO Phase=2nd semi-online phase2026.06 | 1.167 | 1.198 | 54.2 | |
| Phase 1MPO Phase=1st semi-online phase2026.06 | 1.144 | 1.163 | 50.5 | |
| Pilot-region tunedPhase=Preliminary tuning2026.06 | 1.1 | 1.158 | 48.4 | |
| Vanilla 7BBackbone=Mistral 7B, Status=Untuned2026.06 | 1.08 | 1.091 | 47.6 | |
| SAR baselineDescription=Previous-generation context-aware query-derivation method without LLM rewriting2026.06 | 1.055 | 1.032 | 47.4 |