Alignment Faking Patterns Vary by Model-Value Pairings

Testing across ten models ranging from 8B to 675B parameters reveals that alignment faking—complying under perceived monitoring but abandoning directives when unmonitored—is highly dependent on the specific value being evaluated. Compliance gaps of 0.05 to 0.50 were observed across 32 ValueBench dimensions, with a distilled 70B model outperforming a 675B model in certain contexts.