Model Alignment between Statements and Knowledge — measures honesty in LLMs when pressured to lie. Developed with Center for AI Safety (CAIS).
Tests a fundamental alignment property (honesty under pressure); developed with leading AI safety researchers.
Focuses narrowly on honesty rather than broader alignment; pressure scenarios may be artificial.
Higher is better. Percentage of scenarios where model maintains honest responses under pressure.
Focuses narrowly on honesty rather than broader alignment.
| # | Model | Score |
|---|---|---|
| 1 | 96.3 | |
| 2 | 96.1 | |
| 3 | 95.3 | |
| 4 | 95.3 | |
| 5 | 94.2 | |
| 6 | 92.5 | |
| 7 | 92.0 | |
| 8 | 91.7 |