Can Reasoning Models Detect Changes to their Chains of Thought?

This study investigates the ability of recent reasoning models to detect modifications in their chains of thought (CoT), assessing their performance under various conditions, including self and external CoT alterations. The findings reveal that models demonstrate only modest detection accuracy and struggle to identify the nature of the changes, indicating that interventions may not significantly alter model behavior. This research highlights the challenges in ensuring model robustness against tampering, which is crucial for practitioners aiming to improve safety and reliability in AI applications.