This paper, published in Transactions on Machine Learning Research (TMLR) tackles a core challenge in machine learning for sequential systems: a model’s predictions can change the future inputs it later receives, causing errors to compound over time. The paper matters because it proposes a way to close that gap using a continual world model.

A second-year law student reflects on entering legal education as generative AI rapidly reshapes learning, yet remains cautiously acknowledged in academia. Drawing on her engineering background, Maddison contrasts earlier tech-forward training with law school’s more conservative approach, shaped by the school’s new AI policy.