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.

Pablo Bernabeu Pérez
AI safety Researcher, LASR Labs
AI safety researcher at LASR Labs (London AI Safety Research), working in collaboration with Google DeepMind researchers to extend the AI Control framework to chain-of-thought reasoning.
He holds a degree in Computer Science from UPV and a master's in Artificial Intelligence from UPC. Previously, he was a Research Engineer at the Barcelona Supercomputing Center, publishing work on medical language models. He is the co-author of CoT Red-Handed: Stress Testing Chain-of-Thought Monitoring, published at the prestigious NeurIPS conference.