Summarization Preference Learning on TL;DR (test)
87.2Win Rate (0% Flip)α-UDPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| α-UDPOModel=Llama-3.2-3B2026.07 | 87.2 | 86.4 | 78.2 | 86.1 | |
| Dr.DPOModel=Llama-3.2-3B2026.07 | 86.8 | 84.3 | 82.9 | 85 | |
| IPOModel=Llama-3.2-3B2026.07 | 86.5 | 80.3 | 69.3 | 81.8 | |
| UDPOModel=Llama-3.2-3B2026.07 | 86.3 | 85.3 | 77.2 | 85.9 | |
| α-UDPOModel=Qwen-3-1.7B2026.07 | 86.3 | 84.9 | 71.5 | 79.6 | |
| rDPOModel=Llama-3.2-3B2026.07 | 85.9 | 83.5 | 77.3 | 84.7 | |
| UDPOModel=Qwen-3-1.7B2026.07 | 84.6 | 83 | 71.3 | 79.4 | |
| rDPOModel=Qwen-3-1.7B2026.07 | 84.3 | 72.4 | 70.2 | 74.5 | |
| Dr.DPOModel=Qwen-3-1.7B2026.07 | 82.9 | 80.1 | 70.5 | 76.6 | |
| IPOModel=Qwen-3-1.7B2026.07 | 82.1 | 78.9 | 67.1 | 72 | |
| DPOModel=Qwen-3-1.7B2026.07 | 73.9 | 65.4 | 54.5 | 65.9 | |
| DPOModel=Llama-3.2-3B2026.07 | 72.2 | 64.6 | 56.1 | 71.4 | |
| cDPOModel=Llama-3.2-3B2026.07 | 69.6 | 61 | 57 | 70.7 | |
| cDPOModel=Qwen-3-1.7B2026.07 | 67.2 | 60.5 | 51.9 | 63.6 |