Modern sequence models exhibit fragile behavior under distributional changes, delayed feedback, and hidden regime shifts, often failing with high confidence despite degraded internal stability. Existing approaches typically address these failures through static safeguards, post-hoc alignment constraints, or increased model scale, leaving the underlying dynamic instability largely unobserved. We introduce a modular framework for adaptive stability control in sequence models operating under regime uncertainty.
Luis Jaime Ledesma Perez (Fri,) studied this question.