The Future of Facts: Tracing the Factual Generation-Verification Gap

In three linesEmpirical study of the generation-verification gap in LLMs: fact verification is learned before generation, more robust to continual learning, and factual updates create "multi-verse" states where models accept both old and new answers. Analysis across 4 open-source model families at 2 scales.

Summary generated by Claude — human-verified