Ethical alignment evaluation on Open-ended ethical scenarios 20 countries 310 scenarios each
52.1MISVanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaBackbone=Phi-3.5-mini, Decoding Strategy=Vanilla decoding (no DISCA)2026.05 | 52.1 | — | |
| DISCABackbone=Phi-3.5-mini, Decoding Strategy=Full pipeline2026.05 | 51 | 2.13 | |
| VanillaBackbone=Llama-3.3-70B, Decoding Strategy=Vanilla decoding (no DISCA)2026.05 | 47.1 | — | |
| DISCABackbone=Llama-3.3-70B, Decoding Strategy=Full pipeline2026.05 | 43.9 | 6.85 | |
| VanillaBackbone=Phi-4 (14B), Decoding Strategy=Vanilla decoding (no DISCA)2026.05 | 32.4 | — | |
| VanillaBackbone=Qwen2.5-7B, Decoding Strategy=Vanilla decoding (no DISCA)2026.05 | 32.1 | — | |
| DISCABackbone=Qwen2.5-7B, Decoding Strategy=Full pipeline2026.05 | 30.6 | 4.55 | |
| DISCABackbone=Phi-4 (14B), Decoding Strategy=Full pipeline2026.05 | 30.2 | 6.67 |