Personalized Response Generation on AIME Explicit Preference 2025 (test)
69.9PrefTriAlign
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| TriAlign2026.06 | 69.9 | 46.7 | 43.3 | 2.7 | 3.3 | 5.8 | |
| Qwen3-4B-Instruct-25072026.06 | 69.1 | 32.8 | 26.7 | 3.6 | 6.1 | 10.9 | |
| SWIFTYear=20262026.06 | 41.8 | 40 | 30 | 7.2 | 10 | 18 | |
| SFTYear=20222026.06 | 37.5 | 36.1 | 30 | 5.6 | 6.1 | 15.5 | |
| BESTPERSONAYear=20242026.06 | 37.4 | 33 | 26.7 | 5.3 | 6.3 | 16 | |
| P-DEFENSEYear=20252026.06 | 33.9 | 30 | 26.7 | 3.8 | 3.3 | 12.8 | |
| P-DEBIASYear=20242026.06 | 33.1 | 30.9 | 26.7 | 2.5 | 4.2 | 8.2 | |
| 2STEPPROMPTYear=20252026.06 | 18 | 48.2 | 42.3 | 5.1 | 5.9 | 10.6 |