Automated Discovery of Datasets with Surprising Training Side Effects

A new pipeline uses evolution and reinforcement learning to identify datasets that trigger unintended model behaviors, such as persona shifts or misalignment. The method produced 6,575 datasets across eight behaviors that generalize across different model families.