Query Recommendation on Taobao (test)
47.04Cate HR@30AIGQ-Think
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AIGQ-ThinkTraining protocol=IL-SFT+IL-GRPO2026.03 | 47.04 | 66.24 | 7.45 | 9.8 | |
| AIGQ-DirectTraining protocol=IL-SFT+IL-GRPO2026.03 | 46.95 | 63.41 | 6.79 | 6.7 | |
| AIGQ-ThinkTraining protocol=IL-SFT2026.03 | 46.53 | 64.78 | 5.59 | 10.3 | |
| AIGQ-ThinkTraining protocol=SFT+GRPO2026.03 | 44.38 | 64.21 | 6.14 | 8 | |
| AIGQ-ThinkTraining protocol=SFT2026.03 | 44.37 | 63.58 | 5.49 | 7.7 | |
| AIGQ-DirectTraining protocol=IL-SFT2026.03 | 43.05 | 59.46 | 4.42 | 7.5 | |
| AIGQ-DirectTraining protocol=SFT2026.03 | 41.81 | 59.26 | 4.28 | 7.1 | |
| AIGQ-DirectTraining protocol=SFT+GRPO2026.03 | 39.06 | 61.16 | 6.34 | 4.7 | |
| Gemini 3 ProGeneration strategy=zero-shot2026.03 | 34.49 | 54.75 | 0.17 | 7 | |
| GPT-5.1Generation strategy=zero-shot2026.03 | 33.53 | 56.06 | 0.21 | 4.6 | |
| Qwen3-30B-A3BGeneration strategy=zero-shot2026.03 | 30.54 | 55.54 | 0.22 | 5.4 | |
| EBRArchitecture=two-tower DNN2026.03 | 19.98 | 51.98 | 1 | 8.4 |