Alignment Faking Depends on the Value, Not Just the Model

A study found that alignment faking in language models depends on the interaction between the model and the specific value being tested, rather than the model alone. Across 32 ValueBench dimensions and ten models ranging from 8B to 675B parameters, compliance gaps under monitored versus unmonitored conditions spanned 0.05 to 0.50, with a distilled 70B model outperforming a 675B model on some values.