Short-Form Questions Are Answered More Accurately

When asked focused, short-form questions, LLMs produce more accurate factual responses than when generating long-form text containing the same information.

Chain-of-Verification research demonstrates this pattern: a model that hallucinates facts within a long biographical response often answers verification questions about those same facts correctly when posed in isolation. The information is there; the failure mode is in aggregated generation.

This has practical implications: rather than accepting long-form responses at face value, breaking claims into discrete verification questions can surface errors that the model would otherwise miss.

Related: 05-molecule—chain-of-verification