Instruction Alignment on UltraFeedback
23Instruction Following Win (%)MGDA-DECOUPLED
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MGDA-DECOUPLEDModel=Qwen2.5-0.5B-Instruct2026.04 | 23 | 2.1 | 75 | 16.6 | 23 | 60.3 | 19.4 | 4.1 | 76.5 | 21.1 | 6.4 | 72.5 | |
| GROUPDROModel=Qwen2.5-0.5B-Instruct2026.04 | 22.9 | 1.7 | 75.4 | 16.5 | 22.8 | 60.7 | 19.1 | 4.1 | 76.8 | 20.9 | 6.1 | 73 | |
| MGDA-NORMALISEDModel=Qwen2.5-0.5B-Instruct2026.04 | 22.7 | 1.9 | 75.4 | 17.2 | 23 | 59.8 | 19.6 | 4 | 76.4 | 21.4 | 6.2 | 72.4 | |
| CDPOModel=Qwen2.5-0.5B-Instruct2026.04 | 22.3 | 1.7 | 75.9 | 16.6 | 22.6 | 60.8 | 18.6 | 4 | 77.4 | 21 | 6.3 | 72.7 | |
| UNIFORMModel=Qwen2.5-0.5B-Instruct2026.04 | 22.3 | 2 | 75.7 | 16.3 | 22.9 | 60.8 | 19.3 | 4 | 76.7 | 20.9 | 6.5 | 72.6 | |
| REFERENCEModel=Qwen2.5-0.5B-Instruct2026.04 | 20.5 | 1.6 | 77.9 | 16.9 | 21 | 62.1 | 17.4 | 3.8 | 78.8 | 19.6 | 5.9 | 74.5 |