Capabilities Framing Predicts Model Compliance Better Than Safety Framing
August 28, 2026
Research on Qwen3-32B shows that how a model perceives an evaluation affects its compliance. Specifically, capabilities-flavored framing in chain-of-thought predicts compliance with a 24 to 46 percentage-point gap over safety-flavored framing across tested steering conditions.
HOW THIS AFFECTS YOU
●
researcherYou should account for how evaluation framing influences model behavior in safety experiments.
●
policyThis suggests that standard safety evaluations may be sensitive to how tests are worded to the model.