Heading Large language models (LLMs) have achieved remarkable progress in naturallanguage generation, yet they continue to display puzzling behaviors—such asrepetition and incoherence—even when exhibiting low perplexity. Thishighlights a key limitation…