Study identifies a 'generation-verification gap' where language models fail to verify facts they are able to generate

The gap persists across a fact's entire retrieval lifecycle.