Frontier LLMs Show Divergent Responses to Steering Pressure.

The behavior of frontier language models under explicit steering pressure, such as instructions to suppress or elicit reasoning, varies significantly across different models. This research evaluated six leading models using 300 paired base and steered items across three categories. The findings indicate that models differ not only in how much their behavior shifts but also in the *kind* of response they provide, with some response modes being unique to specific models. For instance, GPT-5 was observed to deflect requests to disclose its reasoning while keeping its answer intact, a behavior almost entirely absent in other models. Claude Opus 4.7 and GPT-5 also resisted explicit suppression instructions in distinct ways. The study further traced these behavioral splits to the internal mechanisms of open-weight models like Llama, demonstrating that specific behaviors can be decoded from the residual stream and even injected during generation. These results underscore fundamental differences in how various frontier models are trained and aligned.