Distributional pluralistic alignment on OpinionsQA
1Democrat KL DivergenceCOLD-Kernel
Evaluation Results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| COLD-KernelBackbone=Llama-2-7b-chat-hf2026.03 | 1 | 0.5 | 0.97 | 0.47 | 0.97 | 0.48 | 0.86 | 0.46 | 0.74 | 0.4 | 0.93 | 0.47 | 1.04 | 0.5 | 0.95 | 0.47 | |
| ReFT(vector)Backbone=Llama-2-7b-chat-hf2026.03 | 1.69 | 0.53 | 1.65 | 0.52 | 1.48 | 0.49 | 1.67 | 0.53 | 1.42 | 0.5 | 1.62 | 0.52 | 1.77 | 0.52 | 1.62 | 0.51 | |
| COLD-FDBackbone=Llama-2-7b-chat-hf2026.03 | 2.33 | 0.69 | 2.4 | 0.72 | 1.62 | 0.55 | 2.14 | 0.67 | 2 | 0.62 | 2.26 | 0.7 | 2.32 | 0.7 | 2.17 | 0.7 | |
| BaseBackbone=Llama-2-7b-chat-hf2026.03 | 2.45 | 0.56 | 2.38 | 0.57 | 2.18 | 0.54 | 2.43 | 0.56 | 2.14 | 0.54 | 2.38 | 0.57 | 2.33 | 0.55 | 2.39 | 0.57 |