Knowledge Pull Requests for Continual Document Authoring

The authors introduce Knowledge Pull Requests (KPRs), a framework for continual document authoring that makes updates interpretable. KPRs extract claims, filter and route them to relevant sections, and flag conflicts, generating a ChangeLog that separates claim proposals from document diffs. Evaluated on multilingual Wikipedia revision and RAGTIME query-driven reports, KPRs integrate more information and preserve existing content better than rewriting from sources or regenerating from scratch. Furthermore, KPR-revised articles provide better grounding for question answering than frontier search-enabled models.

KPRs integrate more information and better preserve existing content than rewriting from sources or regenerating from scratch.

KPRs add the most information per token generated.

A KPR-revised article grounds question answering better than a frontier model with search.