Knowledge Pull Requests for Continual Document Authoring
KPRs separate the proposed knowledge change from the text edit that applies it.
The paper frames continual authoring as a traceable update process, not a full rewrite. Its system extracts claims from new material, routes them to document sections, flags conflicts, and emits a ChangeLog plus a document diff. In tests on cross-language Wikipedia revision and RAGTIME report updates, KPRs added more information while preserving more existing content than source-based rewrites or regeneration. The authors also report that a KPR-revised article supported QA better than a frontier search model when the relevant knowledge appeared only in another language. ArXiv · AI/CL/LG's note
The paper frames continual authoring as a traceable update process, not a full rewrite. Its system extracts claims from new material, routes them to document sections, flags conflicts, and emits a ChangeLog plus a document diff. In tests on cross-language Wikipedia revision and RAGTIME report updates, KPRs added more information while preserving more existing content than source-based rewrites or regeneration. The authors also report that a KPR-revised article supported QA better than a frontier search model when the relevant knowledge appeared only in another language. ArXiv · AI/CL/LG's note
score 4