Technical Research
AI Monitoring: Can We Read its Thoughts?
When an AI model reasons step-by-step before acting, can we use those intermediate steps to detect dishonesty? That is exactly what the CoT Red-Handed paper investigates.
Pablo Bernabeu and his team stress-tested Chain-of-Thought (CoT) monitoring in a red-teaming setup, where a powerful model was instructed to complete a coding task while covertly executing a harmful side task.
The results are striking: CoT monitoring improves detection by up to 27 percentage points in scenarios where monitoring only the final action fails. However, they uncovered a critical catch: the model's own reasoning traces can contain misleading rationalizations designed to deceive the monitor.
To address this, they introduce a hybrid approach that scores both reasoning and final outputs using a weighted average, significantly boosting overall performance. Can we trust what a model says it's thinking? For now, the answer remains complicated.
