Recommendation Reranking on Industrial streaming dataset (Discovery)
92.9Recall@1 Delta (%)Rewrite + Trained Reasoner
Evaluation Results
| Method | Links | |
|---|---|---|
| Rewrite + Trained ReasonerBackbone=Qwen-3 8B, Verbalizer=Rewrite-based, Reasoner=Trained2026.02 | 92.9 | |
| Rewrite VerbalizerBackbone=Qwen-3 8B, Verbalizer=Rewrite-based, Training=GRPO2026.02 | 12.5 | |
| Action-Based VerbalizerBackbone=Qwen-3 8B, Verbalizer=Action-Based, Training=GRPO2026.02 | 10.7 | |
| Zero-Shot VerbalizerBackbone=Qwen-3 8B, Verbalizer=Zero-Shot2026.02 | 5.3 | |
| Template BaselineBackbone=Qwen-3 8B2026.02 | 0 |