LLMs Show Persistent Moral Reasoning Gap in Persian Proverb Story Generation

A new dataset PAND pairs Persian proverbs with human-written stories to test abstraction-to-narrative generation; evaluation using LLM-as-a-Judge plus structural metrics finds models achieve fluent output but consistently fail to faithfully encode the underlying moral and causal structure — a gap explicit chain-of-thought reasoning partially closes.