Knowledge Pull Requests for Continual Document Authoring
Knowledge Pull Requests revise documents by routing new claims and logging conflicts instead of full rewrites.
Knowledge Pull Requests support continual document authoring by extracting claims, filtering and routing them to sections, and flagging conflicts with existing content. A ChangeLog separates what knowledge changes from how the text changes. On multilingual Wikipedia revisions and RAGTIME report updates, KPRs integrate more information and preserve existing content better than source rewriting or full regeneration, adding the most information per generated token. A KPR-revised article also grounded question answering better than a frontier model with search, which missed knowledge documented only in other languages.
- KPRs extract, route, and conflict-check new claims
- ChangeLog separates knowledge changes from document diffs
- Beats rewriting and regeneration on Wikipedia and RAGTIME
- KPR articles ground QA better than a searched frontier model
Full article150 words · extracted from huggingface.co · click to collapse
We introduce Knowledge Pull Requests (KPRs), a framework for continual document authoring that makes each change interpretable. Documents require ongoing revision as new knowledge surfaces from other sources, languages, or times, but existing approaches either edit with no account of what knowledge changed or regenerate from scratch. A KPR integrates new knowledge into a document by extracting claims, filtering and routing them to sections, and flagging conflicts with existing content, producing a ChangeLog that separates what knowledge changes (claim proposal) from how the text changes (document diff). We evaluate KPRs on revising Wikipedia across languages and updating query-driven reports on RAGTIME. KPRs integrate more information and better preserve existing content than rewriting from sources or regenerating from scratch, while adding the most information per token generated. A KPR-revised article also grounds question answering better than a frontier model with search, which does not surface knowledge documented only in other languages.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.26634