Personalized Text Generation on News Headline Generation
99.8User-level Accuracy (GT)PARL-0
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| PARL-0Optimization=Scoring function only2026.05 | 99.8 | 99.8 | 99.8 | 99.6 | 99.8 | 99.9 | 99.8 | 99.9 | 0.001 | 100 | |
| PARL-AOptimization=GT-scaled reward (RA)2026.05 | 83.2 | 33.1 | 46.8 | 25.3 | 41.5 | 81.4 | 49.6 | 77.5 | 0.018 | 99 | |
| LM-8BModel=Qwen3-8B2026.05 | 80.4 | 65.6 | 74 | 68.1 | 70.5 | 81.6 | 78.3 | 77.3 | 0.012 | 66.9 | |
| PARL-BOptimization=Margin-only reward (RB)2026.05 | 80.4 | 41.4 | 56.9 | 40.8 | 56.1 | 75 | 59.1 | 74.2 | 0.054 | 100 | |
| LM-235BModel=Qwen3-235B-A22B-Instruct2026.05 | 76.5 | 63.6 | 71.4 | 63.6 | 71.8 | 78.7 | 74.5 | 77.4 | 0.022 | 87.1 |