Frontier AI Models Show Divergent Response Modes Under Steering Pressure

A recent study published on arXiv (2608.06578v1) explores whether advanced language models from various developers display significant differences in behavior when faced with explicit steering pressure. The research assesses six models from different developers, utilizing 300 paired base and steered items across three categories: values-conflict, reasoning-elicitation, and reasoning-suppression, along with 40 validation items. Each model serves as a blind peer judge, evaluating responses according to established behavioral criteria, resulting in a total of 24,480 judgments scored through leave-one-out consensus. The results indicate that the models not only vary in how much steering influences their behavior but also in the types of responses generated, with certain modes appearing in only one or two models. Notably, GPT-5 avoids revealing its reasoning while maintaining its answer (99% vs. 0% for others), and Claude Opus 4.7 shows a distinct response mode. This study underscores how training differences among frontier models contribute to varied behavioral responses under pressure, raising concerns for AI safety and alignment.