Knowledge Pull Requests for Continual Document Authoring

• Ongoing document maintenance must incorporate newly surfaced knowledge, but prompted edits hide what facts change, query-driven systems regenerate from scratch, and raw diffs show words rather than knowledge.Source: §1 · p. 1
• A KPR extracts claims from sources, filters and routes them to document sections, flags conflicts, and produces a ChangeLog separating claim proposals from document diffs.Source: §1 · p. 1
• The study evaluates revision of English Wikipedia articles using other-language counterparts and updates to query-driven reports across temporal, conflict, and balanced RAGTIME variants.Revision adds information about an unchanged world; update synchronizes a document with a changed world.Source: §1 · p. 2
• KPRs are compared with source-conditioned rewriting, rewriting from extracted claims without a claim proposal, and regeneration from all sources; evaluation covers faithfulness, completeness, QA grounding, and change costs.Source: §1 · p. 2
• Across both settings, KPRs integrate more new information and better preserve existing content than baselines; revised Wikipedia articles ground QA better than originals, baselines, or frontier search.The frontier model with web search does not surface knowledge documented only in other languages.Source: §1 · p. 2
• On RAGTIME updates, KPRs flag source conflicts rather than silently resolving them, holding precision when raw-text conditioning is misled; claim conditioning improves completeness, and filtering claims raises precision and recall.KPRs also concentrate changes into contiguous edits a reviewer can approve.Source: §1 · p. 2

Belief Revision and Epistemology of Documents

• KPRs extend belief-revision frameworks by operating on natural-language claims and recording claim proposals separately from text diffs in a ChangeLog.Prior frameworks model belief changes over sentences and logical relations, without representing justifications; KPR experiments cover both revising Wikipedia and updating reports as sources evolve.Source: §2 · p. 2

Knowledge Cutoffs and Editing

• Weight edits and text rewrites can update knowledge, but do not surface which claims changed, conflicts with existing content, or an interpretable update record.Prior work rewrites Wikipedia articles to reflect new sources (Logan IV et al., 2022; Gangi Reddy et al., 2026) and models news revision histories.Source: §2 · p. 2

Claims and Factuality

• Atomic claims support more interpretable factuality assessment and transfer across contexts, motivating KPRs to represent what knowledge each change contributes.FActScore (Min et al., 2023) and VeriScore (Song et al., 2024) verify decomposed subclaims against evidence, while claims also transfer across modalities (Jing et al., 2024; Martin et al., 2026).Source: §2 · p. 3

Main Document and Sources

• A KPR takes a previously written main document and relevant source documents, integrating new information while working with existing content rather than overwriting it.The main can be organized at different granularities, including sections, sentences, paragraphs, or pages.Source: §3 · p. 3

Claims

• KPRs use atomic, decontextualized factual claims from sources and the main to track proposed knowledge, placement, and conflicts precisely (see Gunjal and Durrett, 2024).Non-English sources are decomposed directly into English, while claims from the main are cached offline.Source: §3 · p. 3, Figure 2 · p. 3

ChangeLog

• The ChangeLog separates claim proposals from document diffs, letting reviewers assess added knowledge, inspect textual changes, resolve conflicts, and approve or reject changes before merging.This separation distinguishes a KPR from a simple rewrite and supports collaborative continual authoring.Source: §3 · pp. 3–4

Claim Proposal

• Source claims are classified as absent, conflicting, or supported; absent claims are filtered for relevance and routed to sections, conflicts are flagged, and supported or irrelevant claims are dropped.Candidate claims may also be filtered if the main already covers them or they fail the document’s authoring criterion.Source: §3 · p. 3, Figure 3 · p. 4

4 Method for KPRs

• KPRs use three stages: decompose sources into claims, review and route claims into a proposal, then rewrite affected sections and diff against the original.Coverage and conflict are classified together; relevant absent claims are routed to existing or new sections, while covered, irrelevant, and conflicting claims are dropped or flagged. For non-English sources, claims are decomposed directly into English, and the main document's claims are cached as an index.Source: §4 · p. 4

4.1 Baselines

• ConText rewrites from raw sources, ConClaim rewrites from routed claims without claim review, and Scratch regenerates the document from all sources in one pass.ConText classifies whether a source is relevant to each section; Scratch is used only in RAGTIME.Source: §4.1 · pp. 4–5

4.2 Implementation Details

• All LLM operations use Qwen3.5-27B; experiments impose report outlines for section-based editing, and flagged conflicts are withheld because no human reviewer resolves them.The authors withhold conflicts by default because deciding which source to trust remains an open problem.Source: §4.2 · p. 5

5.2 Grounding Source

• For multilingual QA, the unchanged English article averaged below closed-book, while all rewrites recovered missing knowledge and KPR recovered substantially more than either baseline.The results show that KPR-integrated multilingual information can be used by grounding articles for question answering.Source: Table 1 · p. 6, §5.2 · p. 7

5.1 Article Quality

• Claim conditioning raised InfoP and InfoR-A over raw-text conditioning, and adding the claim proposal raised both again—especially InfoR-A—while retention remained high across methods.KPRs saw the largest benefit in integrating source knowledge.Source: Table 2 · p. 6, §5.1 · p. 6

5.2 Grounding Source

• On 100 hardest multilingual questions, even the smallest 7B model grounded on KPR outperformed the frontier model with web search, which recovered only marginally more answers than the English article.Translating questions to their source language helped closed-book and web-search conditions, but neither approached KPR.Source: Table 4 · p. 7, §5.2 · p. 7
• On English QA, the original article scored highest; rewrites degraded slightly but stayed within half a percent of one another, so KPR's multilingual gains did not meaningfully cost existing-content performance.Smaller models also gained strongly from KPR-revised articles, without requiring a frontier model or search service.Source: Table 3 · p. 6, §5.2 · p. 7

5.3 Edit Cost

• KPR preserved the most text and required the fewest contiguous approvals, integrating more source knowledge than ConText with similar added tokens (2.3× context increase); ConClaim needed over twice KPR's approvals.Without a claim proposal, every claim routed to a section is passed to the rewrite, so the model revises content that did not need to change.Source: Table 5 · p. 7, §5.3 · p. 7

6 Updating Query-Driven Report Generation with KPRs

• The study uses three two-round multilingual RAGTIME variants: temporal arrivals, cross-round conflicts, and balanced nugget splits; report quality is measured with InfoP, InfoR-R, and InfoR-A.ConClaim is omitted because unfiltered RAGTIME claim sets exceed the context window.Source: §6 · p. 8

6.1 Report Quality

• Across all three settings, KPR recalls the most round-1 and round-2 nuggets and leads rewrites in precision; Scratch adds least between rounds, while ConText has lowest precision.The round-1 report itself scores highest on precision because it only needs to be faithful to round-1 sources.Source: §6.1 · pp. 8–9
• Every method recalls more round-1 nuggets than the original report because round-2 sources sometimes repeat information that incremental methods can recover.Incremental methods do not revisit round-1 documents.Source: §6.1 · p. 9
• In the conflict setting, ConText integrates less round-2 information; KPR withholds conflicting claims and adds facts at a comparable rate while achieving the highest precision among methods that integrate meaningful information.ConText must resolve disagreements implicitly during rewriting, which appears to suppress incorporation of incoming information.Source: §6.1 · p. 9

6.2 Edit Cost

• Scratch produces short, high-confidence reports that are most expensive to review; ConText and KPR make similarly sized updates, but ConText adds less information.Only coincidental n-gram overlaps are preserved in Scratch’s generation.Source: §6.2 · p. 9

7 Findings Across Settings

• Claim-based conditioning improves the precision and recall of added information in both experiments, with the same trend in QA, yielding more faithfully grounded documents.The findings extend prior work on claims as a unit for factuality evaluation and retrieval to knowledge integration.Source: §7 · p. 9
• Claim proposals consistently improve rewrite quality over baselines, maintaining precision while raising the recall of integrated information.They also shape rewrites into contiguous edits that are easy to review.Source: §7 · p. 9
• KPRs add the most information per token generated; baselines integrating a comparable volume of text incorporate substantially less source information.Source: §7 · p. 9

• Across cross-lingual Wikipedia revision and query-driven report updating, KPRs integrate information more completely and preserve existing content better than other methods.They also concentrate changes into contiguous edits a reviewer can approve.Source: §8 · p. 9
• KPRs decompose sources into claims, filter and route them, and surface conflicts, producing a ChangeLog that separates knowledge changes from text changes.Source: §8 · p. 9
• By making each change a reviewable claim rather than an opaque edit, KPRs open a path toward collaborative, human-in-the-loop authoring.Source: §8 · p. 9

Computational Efficiency

• Each KPR pipeline stage requires an LLM call, and reducing this per-step compute cost is left to future work.Although KPRs rewrite only changed sections, editcost measures rewrite size and shape rather than the compute used to produce it; lightweight encoders are a possible alternative for classification or routing.Source: Limitations · p. 9

Human Review

• The experiments automate KPRs without evaluating whether human review of ChangeLogs improves editing speed or accuracy.Future work should assess ChangeLogs against text diffs and investigate how humans adjudicate claim proposals; flagged conflicts were withheld from rewriting in these experiments.Source: Limitations · p. 9

A Wikipedia Experiments Implementation Details

• The study samples 599 articles across 49 non-English MegaWika languages and turns article claims into QA pairs, excluding multilingual claims already contained in English Wikipedia.Source: A Wikipedia Experiments Implementation… · p. 12
• QA filtering keeps specific, self-contained factoid questions correctly answered by their gold claims, while Table 8 reports counts for the original, absent/conflict, and final stages.Source: A Wikipedia Experiments Implementation… · p. 13, Table 8 · p. 13
• The 100-question subset samples multilingual questions unanswered by any dense model in closed-book, non-translated settings; translated CB-T and WS-T questions use Qwen3.5-27B.This design targets answers absent from the models’ parametric knowledge so performance must come from the grounding document.Source: A Wikipedia Experiments Implementation… · p. 13
• GPT-5.6 (Sol) is accessed through the OpenAI API at maximum reasoning effort, including in the web-search condition; Table 9 also reports ConText and ConClaim conditioning.Source: A Wikipedia Experiments Implementation… · p. 13

B RAGTIME Experiments Implementation Details

• RAGTIME splits are constructed programmatically from relevance judgments; the temporal split divides relevant documents by publication date, assigning earlier and later halves to rounds 1 and 2.The per-topic cutoff date divides the documents into equal halves.Source: B RAGTIME Experiments Implementation… · p. 13
• The conflict split places documents supporting conflicting OR-nugget answers or complementary AND-nugget information in different rounds, so the conflicting or complementary document appears in round 2.Source: B RAGTIME Experiments Implementation… · p. 13
• The balanced split partitions documents to make the number of nuggets in each round as equal as possible; Table 10 reports novel nugget counts per round.Source: B RAGTIME Experiments Implementation… · p. 13

C Method Details

• Cross-lingual decomposition matches native-then-translate in faithfulness and is chosen over translate-then-decompose because it needs one LLM pass rather than two.The compared strategies translate before decomposing, decompose before translating claims, or decompose source sentences directly into English claims.Source: C Method Details · p. 13

D Metric Details

• MiRAGE (Martin et al., 2026) scores predicted text against reference evidence using atomic subclaims and a cross-lingual Qwen3.5-27B support judge, reporting InfoP and InfoR variants.The reported recall variants are InfoR-A and InfoR-R.Source: §D.1 · p. 14

D.1.1 Information Precision

• InfoP is a source-constrained FActScore (Min et al., 2023) variant that scores predicted-text subclaims against reference documents, using added Wikipedia sentences or the round-2 RAGTIME report as predictions.For Wikipedia, reference evidence is the non-English sources; for RAGTIME, it is the union of round-1 and round-2 source documents.Source: §D.1.1 · p. 14

D.1.2 Information Recall

• InfoR reverses the precision comparison, with InfoR-R measuring preserved original or round-1 claims and InfoR-A measuring added non-English-source or round-2 claims.For Wikipedia, the retained claims come from the original English article and added claims from non-English sources; for RAGTIME, they come from the round-1 and round-2 nuggets, respectively.Source: §D.1.2 · p. 14

D.2 Edit Cost

• WER measures word-level edit distance, while Click counts approvals for each contiguous inserted or replaced span, regardless of span length.Click is reported as a per-document mean, so many scattered edits can require more clicks than a few large blocks.Source: §D.2 · p. 15
• Tok counts added word tokens, Presv measures verbatim retention, and Add measures rewritten-to-original size; neither Presv nor Add has a fixed preferred direction.A sound revision may rewrite much of a document; Wikipedia compares the English article with the reconstructed rewrite, while RAGTIME compares its round-1 seed report with the round-2 report.Source: §D.2 · p. 15