LLM Alignment on UltraFeedback (in-domain)
80.6Win Rate (KL, alpha=1)GEB-π
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GEB-πBase Model=Mistral-Instruct-v0.32025.09 | 80.6 | 0.7531 | 74.41 | 0.6723 | 51.75 | 0.5986 | |
| GEB-1/πBase Model=Mistral-Instruct-v0.32025.09 | 79.69 | 0.7614 | 73.87 | 0.6342 | 51.86 | 0.6088 | |
| GEB-arctanh(π − 1)Base Model=Mistral-Instruct-v0.32025.09 | 79.6 | 0.7544 | 74.57 | 0.6642 | 51.69 | 0.5962 | |
| f-DPOBase Model=Mistral-Instruct-v0.32025.09 | 77.49 | 0.7426 | 72.7 | 0.6488 | 50.83 | 0.5902 | |
| GEB-arctanh(π − 1)Base Model=LLaMA-3-8B-SFT2025.09 | 74.55 | 0.8032 | 72.57 | 0.7902 | 68.34 | 0.7612 | |
| GEB-1/πBase Model=LLaMA-3-8B-SFT2025.09 | 74.36 | 0.8084 | 72.28 | 0.7733 | 67.99 | 0.7618 | |
| GEB-πBase Model=LLaMA-3-8B-SFT2025.09 | 73.9 | 0.8048 | 72.08 | 0.7921 | 67.79 | 0.7698 | |
| f-DPOBase Model=LLaMA-3-8B-SFT2025.09 | 72.89 | 0.7926 | 71.19 | 0.7915 | 66.62 | 0.7566 |