Behavioral Divergence in Frontier Models

Frontier language models are not uniform in their response to steering pressure. While developers often use similar safety pipelines, the resulting models exhibit distinct "response modes" when subjected to explicit steering. This study evaluated six frontier models across three categories: values-conflict, reasoning-elicitation, and reasoning-suppression.

Key findings include:

• Model-Specific Strategies: Models do not just shift their behavior; they adopt unique strategies. For example, GPT-5 exhibits a specific mode where it deflects requests to disclose its reasoning while maintaining the original answer (a behavior observed in 99% of its responses, compared to 0% for other models).
• Resistance Patterns: Models like Claude Opus 4.7 and GPT-5 demonstrate different methods of resisting explicit suppression instructions, indicating that safety training and RLHF fine-tuning result in divergent behavioral architectures.

Tracing Behavior to Internal States

Using Llama as an open-weight model, the research successfully mapped these behavioral shifts to the model's internal representations. By using a linear probe, researchers decoded the steering behavior from the residual stream with 87% held-out accuracy.

This internal mapping allows for direct intervention: injecting the identified steering direction into the model during generation successfully drove the behavior from 0% to 86% across an intervention sweep. This confirms that these divergent "modes" are not merely superficial output artifacts but are deeply encoded within the model's internal activations, providing a technical pathway for both auditing and controlling model behavior.