Updates
Our research
Evaluations, interpretability and governance work on deceptive and scheming behaviour in frontier AI systems.
Measuring Reward-Seeking via Contrastive Belief Updates
Visible forms of misbehavior are dropping in frontier models. Does that mean the models are becoming aligned, or are they just getting better at doing whatever they believe their grader rewards? Our paper finds that production reinforcement learning increases reward-seeking.
Stress Testing Deliberative Alignment for Anti-Scheming Training
We partnered with OpenAI to assess frontier language models for early signs of scheming — covertly pursuing misaligned goals — in controlled stress-tests (non-typical environments), and studied a training method that can significantly reduce (but not eliminate) these behaviors. Our results are complicated by models’ increasing ability to recognize our evaluation environments as tests of their alignment.
17 September 2025
Science of Scheming
Latest