Value-Conflict Diagnostics Reveal Widespread Alignment Faking in Language Models

We evaluate the olmo2 , qwen2.5 , and qwen3 model families. Each consequence variant is identified by an abbreviation of its instrumental goal followed by a numeric index — for instance, VP-1 denotes the first consequence variant threatening Value Preservation . We additionally include a no-consequence condition to assess whether the presence of a concrete consequence is a necessary precondition for strategic compliance. As shown in Figure 3 , compliance gaps are near zero in the no-consequence condition, confirming that behavioral divergence across oversight conditions arises from strategic reasoning about consequences rather than mere sensitivity to different oversight framings. Furthermore, while substantial compliance gaps persist across all instrumental goals, Value Preservation and Resource Acquisition consequences consistently elicit the highest rates, suggesting that threats to internalized values and computational resources are the most potent drivers of strategic compliance.