A framework that logs which claims changed when a document is revised

single source· 1 articles · confidence: medium · first seen 2026-09-21 20:00 UTC

What this means for you

Nothing to act on today: no code release, no model, no API, and the results are the authors' own on two benchmarks with no evaluation date given. If you build document-updating or retrieval-augmented writing tools, the transferable idea is keeping the knowledge change separate from the text diff.

Researchers propose Knowledge Pull Requests, a way to revise a document that records what knowledge changed, not just how the text changed. New material is split into claims, routed to sections, and checked against existing text for conflicts; the output is a ChangeLog separating claim changes from the document diff. Tests were on cross-language Wikipedia revisions and on query-driven reports in RAGTIME. The authors report more information added per token generated and better preservation of existing content than rewriting from sources or regenerating from scratch, and say the revised articles answer questions better than a frontier model with search. No evaluation date or harness is given.

Key facts

  • ·The framework is called Knowledge Pull Requests, described in arXiv paper 2609.26634, posted 21 September 2026. source
  • ·Output is a ChangeLog that separates claim proposals (what knowledge changed) from the document diff (how the text changed). source
  • ·Evaluated on revising Wikipedia articles across languages and on updating query-driven reports in RAGTIME. source
  • ·The authors report the method adds the most information per token generated and preserves existing content better than rewriting from sources or regenerating from scratch. source
  • ·The authors report a KPR-revised article grounds question answering better than a frontier model with search, which they say misses knowledge documented only in other languages. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire