Researchers have developed a method for controlling AI policies in high-stakes environments. The approach, called Conformal Policy Control, uses a safe reference policy to regulate the behavior of an optimized but untested policy. This allows for exploration and improvement without violating safety constraints. The theory provides finite-sample guarantees even for non-monotonic loss functions, making it more robust than previous methods. Experiments on various applications show that this approach can improve performance while ensuring safety. --- Why it matters: This matters to AI researchers because it addresses the trade-off between exploration and safety in high-stakes environments. By providing a framework for controlling policy behavior, Conformal Policy Control enables safer and more efficient deployment of AI systems. Source: https://arxiv.org/abs/2603.02196