Preference Alignment on PRISM normalized-step (test)
2.328Borda AvgHard Panel
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Hard PanelBackbone=Llama-3.2-1B, Bootstrap=n=10002026.02 | 2.328 | 9,836.294 | |
| US-RepBackbone=Llama-3.2-1B, Bootstrap=n=10002026.02 | 2.171 | 5,125.712 | |
| Soft PanelBackbone=Llama-3.2-1B, Bootstrap=n=10002026.02 | 1.88 | -3,600.814 | |
| Full PRISMBackbone=Llama-3.2-1B, Bootstrap=n=10002026.02 | 1.842 | -4,729.654 | |
| BaseBackbone=Llama-3.2-1B, Bootstrap=n=10002026.02 | 1.779 | -6,631.538 |