Pith. sign in

REVIEW 5 major objections 4 minor 2 cited by

ScholaWrite: A Dataset of End-to-End Scholarly Writing Process

T0 review · 5 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The first dataset to record scholarly writing keystroke-by-keystroke, with each edit labeled by the cognitive intention behind it.

desk verdict First longitudinal keystroke-level scholarly writing dataset with intention annotations; the capture method's fidelity is the main risk and the data isn't released yet, but the descriptive analyses are careful and the limitations are honestly stated. read the letter →

arxiv 2502.02904 v5 pith:3LZCSWS4 submitted 2025-02-05 cs.HC cs.CLq-bio.NC

Khanh Chi Le , Linghe Wang , Minhwa Lee , Ross Volkov , Luan Tuyen Chau , Dongyeop Kang This is my paper · ORCID
classification cs.HCcs.CLq-bio.NC
keywords scholarlywritingprocesskeystrokeloggingintentiontaxonomyOverleafcognitiveannotationLLMassistanceanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the months-long cognitive process behind a finished scientific manuscript can be captured empirically, and it presents the first dataset that does so: ScholaWrite. A Chrome extension logged keystrokes in Overleaf as ten graduate students wrote five computer science preprints over four months, yielding nearly 62,000 text changes, each annotated with one of fifteen writing intentions. The data show that scholarly writing is non-linear and multi-intentional, with most sessions interleaving three or more intentions and writing time shifting from planning toward revision as manuscripts mature. The same data are used to show that current large language models imitate surface-level edits but cannot reliably predict a writer's next intention from text alone. If the dataset holds up, it gives the field a window into writing-as-thinking and a training resource for assistants that anticipate a writer's next move rather than merely completing text.

What carries the argument

The load-bearing machinery has two parts. A Chrome extension captures the visible text of the Overleaf editor on every key-up event and diffs successive snapshots with the diff_match_patch algorithm, producing a per-keystroke record of edits with timestamp, file name, and author ID. A taxonomy of fifteen cognitive writing intentions — grouped into Planning (idea generation, idea organization, section planning), Implementation (text production, object insertion, citation integration, cross-reference, macro insertion), and Revision (fluency, coherence, clarity, structural, linguistic style, scientific accuracy, visual formatting) — converts those raw diffs into spans of cognitive activity. Two segmentation units, intention sessions and sessions, defined by a 10-minute inactivity cutoff with minimum lengths of 30 seconds and one minute respectively, carry the analyses of how intentions mix, transition, and shift across the writing lifecycle.

What would settle it

Run the extension on a fresh writing session while also capturing full-document state on every change, for example through Overleaf's own version history or a shadow snapshot of the whole project, then check whether the diffed keystroke stream reconstructs the true edit sequence; any substantial divergence — edits executed during scrolls, fast bursts, or in non-visible files — would show the capture is incomplete and undercut the dataset's fidelity claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that ScholaWrite is the first dataset to capture the end-to-end scholarly writing process, from initial drafts to final manuscripts, through in-situ keystroke logging and cognitive annotation. Every recorded edit links the before and after text of a keystroke-level change to one of fifteen mutually exclusive writing intentions, grouped into Planning, Implementation, and Revision, so the corpus records not only what changed in a manuscript but why. From this, the paper argues that scholarly writing is recursive and phase-dependent: more than half of all writing sessions involve three or more intertwined intentions, and the share of time devoted to revision grows sharply in the final third of a project. It further claims that current LLMs are misaligned with this process — base models score at most 0.12 weighted F1 at predicting the next intention from text alone, fine-tuning on ScholaWrite lifts this to 0.64 for BERT-class models, and the strongest evaluated models align best with revision intents while degrading on planning and implementation. The dataset is offered as a lens on the cognitive dynamics of writing, a benchmark for evaluating writing assistance, and a training resource for cognitively aligned assistants.

Load-bearing premise

The load-bearing premise is that key-up-based sampling of the visible Overleaf editor text, diffed with diff_match_patch, records the complete sequence of meaningful writing actions; if edits made during scrolling, rapid typing, or outside the visible pane are missed or distorted, every downstream statistic and model evaluation inherits that bias.

Editorial extensions

If this is right

  • Fine-tuning on ScholaWrite lifts next-intention prediction from a weighted F1 of at most 0.12 to 0.64 for BERT-class models, so cognitively annotated keystroke data can teach models to anticipate a writer's next move.
  • The alignment experiments quantify a concrete gap: current LLMs match human edits best on revision intentions such as clarity and fluency, and degrade on planning and implementation, so assistants aimed at the heavy cognitive work will need a different training signal.
  • The dataset supplies a benchmark for evaluating writing assistance against the writing process itself, not just against the final text.
  • The Chrome extension plus the annotation pipeline is offered as a reusable, ethically approved method for collecting realistic longitudinal writing data in Overleaf at scale.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: the before-text / after-text / intention triplets could be reassembled into a training regime for models that plan, draft, and revise iteratively, since the dataset records the edit and its cause rather than only the finished sentence; the paper builds a first version of this in its iterative self-writing experiment but does not develop it fully.
  • My inference: the segmentation rules themselves may shape the headline results — the 10-minute inactivity cutoff and the 30-second minimum discard short bursts, so the 'planning is rare' finding should be re-tested with coarser and finer session boundaries.
  • My inference: the taxonomy's universality is testable beyond computer science; applying the same fifteen labels to humanities or biology writing, where tools and conventions differ, would show whether the intentions are general cognitive categories or specific to LaTeX-driven CS paper writing.
  • My inference: because only one of the five projects is multi-author, the dataset cannot yet speak to collaborative cognition; a direct extension would run the same extension in team projects and study how the intention streams of co-authors interleave.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The manuscript introduces ScholaWrite, described as the first dataset capturing the end-to-end scholarly writing process through a custom Chrome extension that logs keystroke-level edits on Overleaf. The dataset comprises 61,504 annotated writing actions from five computer science preprints, labeled with a 15-category taxonomy of cognitive writing intentions. The paper also presents descriptive analyses of time allocation, session duration, transition patterns, and early/late-phase shifts, and evaluates LLMs on next-intention prediction and output alignment. The advertised contributions are the collection tool, the annotated corpus, and empirical insights into human-LLM gaps.

Significance. If the dataset is released and the capture pipeline is validated, ScholaWrite would be a genuinely novel resource: no comparable public corpus traces the full multi-month writing process of scientific manuscripts at keystroke granularity with intent annotations. The paper documents IRB approval, privacy filtering, annotation procedures, and prompt templates in detail, which aids reproducibility. The main risks are that the central artifact is not yet publicly accessible, the capture method is not validated against a ground truth, and the paper's own human evaluation partially contradicts the claim that ScholaWrite improves LLM writing behavior.

major comments (5)
  1. [§3.1, Appendix A.2] The 'end-to-end' claim rests on the key-up/visible-text diffing scheme, but the paper provides no evidence that this scheme reconstructs the complete edit sequence of a document. Overleaf's editor is viewport-rendered; typing while scrolled, mouse-only insertions, drag-and-drop, autocomplete, and external synchronization can be missed or represented as large spurious diffs. The post-hoc filters (artifact removal, diff-length cap of 300 in Appendix A.3) do not address this and may discard genuine large edits such as pasted paragraphs. Since all statistics in Tables 3–5 and the model evaluations in Section 6 are computed from these diffs, this is load-bearing. Please add a validation study replaying the recorded diffs and checking reconstruction of the final arXiv source (or another ground truth), and report the proportion of events dropped by each filter.
  2. [§3.3, Appendix B] The reported inter-annotator agreement of 0.71 was computed on a 1K-keystroke subset annotated by the same two authors who developed the taxonomy, after iterative discussion with a cognitive linguist. This is not an independent reliability estimate and likely overstates agreement for a third party applying the codebook. Please report per-label agreement, use at least one independent annotator, or explicitly present the 0.71 as a design-internal consistency measure rather than a validation.
  3. [§6 and Appendix G.7] The paper's framing that 'current LLMs cannot infer writing intent' and that fine-tuning on ScholaWrite improves performance 'dramatically' is not supported by its own numbers. Table 6 shows Llama-8B-SW at 0.13 weighted F1, essentially unchanged from the 0.12 base, while the large gains are for BERT/RoBERTa. More seriously, the human evaluation in Tables 11–14 shows Llama-8B-SW underperforming Llama-8B-Zero on accuracy, fluency, coherence, and relevance across all four seeds, and the appendix itself states the fine-tuned model 'did not outperform the baseline.' The main text should either temper the claims to the specific models and tasks where gains hold, or explain the discrepancy; as written, the 'human-LLM gaps' contribution is overstated.
  4. [§4.1 and §7] The paper repeatedly calls ScholaWrite the 'first large-scale publicly available dataset,' but no URL, repository, or sample is provided anywhere in the manuscript. For a dataset paper, the artifact itself is the central contribution; without at least a hosted version, the descriptive analyses cannot be checked or reused. Please release the dataset (or an anonymized sample and schema) with the final version, or clearly label the paper as a description of an upcoming release.
  5. [§4.2, Tables 3–5] The session definitions depend on ad hoc thresholds: 10-minute inactivity, 30-second minimum intention-session duration, and 1-minute minimum session duration. These thresholds directly determine the counts in Tables 4–5 and the transition matrix in Figure 2, but no sensitivity analysis is reported. A short robustness check (e.g., 5/15-minute inactivity cutoffs, 15/60-second minimum durations) would show whether the qualitative findings are stable or artifacts of the segmentation.
minor comments (4)
  1. [Throughout] There are several copy-editing issues: 'LATEX' appears with literal LaTeX markup in the abstract, 'SCHOLAWRITEunderscores' lacks a space, and some figure captions (e.g., Figure 7) contain stray text. A final proofreading pass is needed.
  2. [Table 6] The GPT-4o row has an empty '+SW' cell; since GPT-4o was not fine-tuned, mark this as N/A and explain why in the caption.
  3. [Figure 2] The transition matrix uses cryptic abbreviations (I.G, S.P, C.I, C, etc.); the caption should spell out all labels or use the full names in the figure itself to make the data interpretable.
  4. [Appendix G.7.3] The statement that 'no LLMs were used during any stage of the study, except for grammatical error correction in this manuscript' is confusing given the paper's own LLM experiments; please rephrase to clarify the intended scope (e.g., data annotation or analysis).

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; only a minor self-citation in the taxonomy background that does not drive the main results.

full rationale

This is a dataset-and-analysis paper, not a derivation paper: the central contributions are a keystroke-capture tool, a 61K-action annotated corpus, descriptive statistics over the annotations, and supervised LLM benchmarks on held-out portions of that corpus. No equation in the paper reduces a prediction to a fitted input. The one self-citation overlap is the taxonomy background: Section 2 says 'Koo et al. (2023) proposed a taxonomy of scholarly writing based on keystroke data from short, 30-minute research-plan sessions,' and Section 3.3 says the annotation taxonomy was built 'Building on cognitive theories of writing (Flower and Hayes, 1981) and recent empirical studies on scholarly revision (Du et al., 2022b; Koo et al., 2023).' Koo et al. shares authors with this paper, so this is a genuine self-citation, but it is not load-bearing: the 15-intention taxonomy was re-derived by iterative open coding with a reported weighted F1 of 0.71 in Appendix B.2, and the paper's headline findings, transition statistics, phase comparisons, and LLM evaluations rest on the collected and annotated data rather than on the prior taxonomy's conclusions. The LLM next-intention and alignment experiments use the human annotations as labels for standard supervised fine-tuning and held-out test evaluation, which is benchmark practice rather than circular reasoning. The capture-fidelity concern about key-up and visible-text sampling is a data-quality risk, not a circularity, because the paper's statistics would be biased but would not be logically forced by the input definitions.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the observational fidelity of the capture system and the reliability of the two author-annotators' cognitive labels. The free parameters listed are segmentation thresholds chosen by the authors; they are not fitted to produce a specific outcome, but they do shape the descriptive results. No external entity is required beyond the taxonomy categories.

free parameters (6)
  • Session inactivity threshold = 10 minutes
    Ends a session after 10 minutes of no typed activity. Chosen by the authors, not derived from data. All session-level statistics (e.g., 57% multi-intent sessions) depend on this cutoff.
  • Minimum intention session duration = 30 seconds
    Discards intention sessions shorter than 30 seconds as trivial. Influences the duration and count statistics in Table 4.
  • Minimum session duration = 1 minute
    Overall sessions must last at least one minute. Affects session-level analyses.
  • Early/late phase split = first third / final third of project duration
    Defines early and late phases. Changes in Table 5 and Figure 4 depend on this arbitrary split.
  • Diff length cap for model training = 300
    Keystroke entries with a diff array larger than 300 are excluded from LLM training. Could bias the training set toward smaller edits.
  • IAA subset size = 1,000 keystrokes
    Inter-annotator agreement (weighted F1=0.71) was computed on only 1,000 of the 61,504 entries, a small reliability check.
assumptions (3)
  • domain assumption Flower and Hayes (1981) cognitive process theory of writing (planning, translating, reviewing) provides a valid high-level model of writing.
    The taxonomy is built on this theory. The paper does not challenge its validity. Invoked in Sections 2 and 3.3.
  • domain assumption Keystroke logging is a valid and unobtrusive method for observing writing processes.
    Taken from prior work (Leijten and Van Waes 2013). The entire dataset collection rests on this premise. Invoked in Section 2.
  • domain assumption The 15-intention taxonomy is mutually exclusive and collectively exhaustive.
    Claimed by the authors based on iterative coding of a subset, with a weighted F1 of 0.71 on 1,000 entries. Full-corpus validation is not reported. Invoked in Section 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ScholaWrite: A Dataset of End-to-End Scholarly Writing Process." pith.science (2026). https://pith.science/paper/3LZCSWS4

@misc{pith2026250202904,
  author       = {Pith},
  title        = {Pith review of: ScholaWrite: A Dataset of End-to-End Scholarly Writing Process},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3LZCSWS4}},
  note         = {Machine review of arXiv:2502.02904}
}
read the original abstract

Writing is a cognitively demanding activity that requires constant decision-making, heavy reliance on working memory, and frequent shifts between tasks of different goals. To build writing assistants that truly align with writers' cognition, we must capture and decode the complete thought process behind how writers transform ideas into final texts. We present ScholaWrite, the first dataset of end-to-end scholarly writing, tracing the multi-month journey from initial drafts to final manuscripts. We contribute three key advances: (1) a Chrome extension that unobtrusively records keystrokes on Overleaf, enabling the collection of realistic, in-situ writing data; (2) a novel corpus of full scholarly manuscripts, enriched with fine-grained annotations of cognitive writing intentions. The dataset includes \LaTeX-based edits from five computer science preprints, capturing nearly 62K text changes over four months; and (3) analyses and insights into the micro-dynamics of scholarly writing, highlighting gaps between human writing processes and the current capabilities of large language models (LLMs) in providing meaningful assistance. ScholaWrite underscores the value of capturing end-to-end writing data to develop future writing assistants that support, not replace, the cognitive work of scientists.

Figures

Figures reproduced from arXiv: 2502.02904 by the authors.

Figure 1
Figure 1. An example scholarly writing process with an [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Transition probability matrix between writing [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The number of intentions per writing session [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (22 more)
Figure 4
Figure 4. Figure 4: Dynamics across early and late phases of writing. (a) The share of time devoted to each intention shifts [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Model alignment patterns. (a; minutes-vs￾alignment) Longer writing sessions show lower align￾ment, indicating higher cognitive complexity. (b; #- intents-vs-alignment) Alignment decreases as more in￾tentions intertwine within a session. 5 10 duration 0.75 0.80 0.85 0.9…
Figure 7
Figure 7. Figure 7: The array of differences between two subse [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The Chrome extension interface (A) on the [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Annotation interface. During the annotation [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Wasserstein distance to uniform distribution for each distribution of writing intentions. Orange, Blue, [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Distribution of labels over time across projects. Orange, Blue and Purple represent Planning, Implemen [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Distribution of high-level intention activities over time. Orange, Blue and Purple represent Planning, [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Distribution of Per-intention writing activities over time. Orange, Blue and Purple represent Planning, [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Sankey diagrams representing the intention flow of each project. Figure (a) to (d) generated from all [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Transition probabilities across time. (a) Early sessions vs. (b) Late sessions. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 17
Figure 17. Figure 17: Writing duration (minutes) vs. model alignment scores (Levenshtein ratio and BERTScore-F1) across [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]
Figure 18
Figure 18. Figure 18: Number of writing intentions vs. alignment scores (Levenshtein ratio and BERTScore-F1) for GPT-5 [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]
Figure 19
Figure 19. Figure 19: The overview of next writing intention pre [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]
Figure 20
Figure 20. Figure 20: Metric scores of the final writing output of models after 100 iterations of the iterative self-writing [PITH_FULL_IMAGE:figures/full_fig_p023_20.png]
Figure 21
Figure 21. Figure 21: Human evaluation results of iterative writing outputs of models [PITH_FULL_IMAGE:figures/full_fig_p023_21.png]
Figure 22
Figure 22. Figure 22: Distribution of high-level writing activities on seed 3 over time by models. [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]
Figure 23
Figure 23. Figure 23: Experiment Setup for the iterative self-writing [PITH_FULL_IMAGE:figures/full_fig_p024_23.png]
Figure 24
Figure 24. Figure 24: presents the sample outputs from the fine￾tuned Llama-8B model during the iterative self￾writing experiment. The model successfully was able to add several LaTeX commands to put some custom icon images (“Macro Insertion”). Also, it successfully revised several words a…
Figure 25
Figure 25. Figure 25: Distribution of high-level writing activities over time by models - [PITH_FULL_IMAGE:figures/full_fig_p031_25.png]
Figure 26
Figure 26. Figure 26: Distribution of Per-intention writing activities over time by models - [PITH_FULL_IMAGE:figures/full_fig_p032_26.png]
Figure 27
Figure 27. Figure 27: The user interface for the human evaluation process. [PITH_FULL_IMAGE:figures/full_fig_p033_27.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Privacy-Preserving Proof of Human Authorship via Zero-Knowledge Process Attestation

    cs.CR 2026-02 unverdicted novelty 6.0 of 10

    ZK-PoP uses zero-knowledge proofs to attest that a writing session matches human typing patterns, without exposing the underlying behavioral data.

  2. Detecting Cognitive Signatures in Typing Behavior for Non-Intrusive Authorship Verification

    cs.CR 2026-02 unverdicted novelty 6.0 of 10

    Cognitive Load Correlation from keystroke timings distinguishes genuine human composition from mechanical transcription with estimated 85-95% accuracy in a non-intrusive framework.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    Samir Bourekkache. 2022. English for specific purposes: writing scientific research papers. case study: Phd students in the computer science department

    work page 2022

  2. [2]

    Sathena Chan. 2017. Using keystroke logging to understand writers’ processes on a reading-into-writing test. Language Testing in Asia, 7:1--27

    work page 2017

  3. [3]

    Mike D ' Arcy, Alexis Ross, Erin Bransom, Bailey Kuehl, Jonathan Bragg, Tom Hope, and Doug Downey. 2024. https://doi.org/10.18653/v1/2024.acl-long.377 ARIES : A corpus of scientific paper edits made in response to peer reviews . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6985--...

    doi:10.18653/v1/2024.acl-long.377 2024

  4. [4]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36

    2024

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171--4186

    2019

  6. [6]

    Paul B Diederich. 1974. Measuring growth in english

    work page 1974

  7. [7]

    Wanyu Du, Zae Myung Kim, Vipul Raheja, Dhruv Kumar, and Dongyeop Kang. 2022 a . https://doi.org/10.18653/v1/2022.in2writing-1.14 Read, revise, repeat: A system demonstration for human-in-the-loop iterative text revision . In Proceedings of the First Workshop on Intelligent and Interactive Writing Assistants (In2Writing 2022), pages 96--108, Dublin, Irelan...

    work page doi:10.18653/v1/2022.in2writing-1.14 2022

  8. [8]

    Wanyu Du, Vipul Raheja, Dhruv Kumar, Zae Myung Kim, Melissa Lopez, and Dongyeop Kang. 2022 b . https://doi.org/10.18653/v1/2022.acl-long.250 Understanding iterative revision from human-written text . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3573--3590, Dublin, Ireland. Associ...

    doi:10.18653/v1/2022.acl-long.250 2022

Show all 40 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  2. [10]

    Julen Etxaniz, Oscar Sainz, Naiara Miguel, Itziar Aldabe, German Rigau, Eneko Agirre, Aitor Ormazabal, Mikel Artetxe, and Aitor Soroa. 2024. https://doi.org/10.18653/v1/2024.acl-long.799 Latxa: An open language model and evaluation suite for B asque . In Proceedings of the 62n...

    2024 doi

  3. [11]

    Linda Flower and John R. Hayes. 1981. http://www.jstor.org/stable/356600 A cognitive process theory of writing . College Composition and Communication, 32(4):365--387

    1981

  4. [12]

    Hongwen Guo, Paul D Deane, Peter W van Rijn, Mo Zhang, and Randy E Bennett. 2018. Modeling basic writing processes from keystroke logs. Journal of Educational Measurement, 55(2):194--216

    2018

  5. [13]

    Takumi Ito, Tatsuki Kuribayashi, Hayato Kobayashi, Ana Brassard, Masato Hagiwara, Jun Suzuki, and Kentaro Inui. 2019. Diamonds in the rough: Generating fluent sentences from early-stage drafts for academic writing assistance. arXiv preprint arXiv:1910.09180

    2019 arXiv

  6. [14]

    Chao Jiang, Wei Xu, and Samuel Stevens. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.641 ar X iv E dits: Understanding the human revision process in scientific writing . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9420-...

    2022 doi

  7. [15]

    Roger Johansson, sa Wengelin, Victoria Johansson, and Kenneth Holmqvist. 2010. Looking at the keyboard or the monitor: relationship with text production processes. Reading and writing, 23:835--851

    2010

  8. [16]

    L \'e ane Jourdan, Florian Boudin, Richard Dufour, and Nicolas Hernandez. 2023. Text revision in scientific writing assistance: An overview. arXiv preprint arXiv:2303.16726

    2023 arXiv

  9. [17]

    L \'e ane Jourdan, Florian Boudin, Nicolas Hernandez, and Richard Dufour. 2024. Casimir: A corpus of scientific articles enhanced with multiple author-integrated revisions. arXiv preprint arXiv:2403.00241

    2024 arXiv

  10. [18]

    Elena D Kallestinova. 2011. How to write your first research paper. The Yale journal of biology and medicine, 84(3):181

    2011

  11. [19]

    Keita Kobayashi, Kohei Koyama, Hiromi Narimatsu, and Yasuhiro Minami. 2022. https://aclanthology.org/2022.lrec-1.609 Dataset construction for scientific-document writing support by extracting related work section and citations from PDF papers . In Proceedings of the Thirteenth...

    2022

  12. [20]

    Ryan Koo, Anna Martin, Linghe Wang, and Dongyeop Kang. 2023. Decoding the end-to-end writing trajectory in scholarly manuscripts. arXiv preprint arXiv:2304.00121

    2023 arXiv

  13. [21]

    Alexandra Rowe Krapels. 1990. https://api.semanticscholar.org/CorpusID:60659113 Second language writing: An overview of second language writing process research

    1990

  14. [22]

    a nder, Daniel Rau, Maximilian R \

    Dennis Kundisch, Jan Muntermann, Anna Maria Oberl \"a nder, Daniel Rau, Maximilian R \"o glinger, Thorsten Schoormann, and Daniel Szopinski. 2021. An update for taxonomy designers: methodological guidance from information systems research. Business & Information Systems Engine...

    2021

  15. [23]

    Ilia Kuznetsov, Jan Buchmann, Max Eichler, and Iryna Gurevych. 2022. Revise and resubmit: An intertextual model of text-based collaboration in peer review. Computational Linguistics, 48(4):949--986

    2022

  16. [24]

    Mari \"e lle Leijten and Luuk Van Waes. 2013. Keystroke logging in writing research: Using inputlog to analyze and visualize writing processes. Written Communication, 30(3):358--392

    2013

  17. [25]

    Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Yi Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Scott Smith, Yian Yin, and 1 others. 2024. Can large language models provide useful feedback on research papers? a large-scale empirical analysis. NEJM AI, 1(8):...

    2024

  18. [26]

    Eva Lindgren and Kirk Sullivan. 2019. Observing writing: Insights from keystroke logging and handwriting, volume 38. Brill

    2019

  19. [27]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://api.semanticscholar.org/CorpusID:198953378 Roberta: A robustly optimized bert pretraining approach . ArXiv, abs/1907.11692

    2019 arXiv

  20. [28]

    Liang Lu, Peirong Xie, and David Mortensen. 2024. https://doi.org/10.18653/v1/2024.acl-long.788 Semisupervised neural proto-language reconstruction . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14715...

    2024 doi

  21. [29]

    Charles A MacArthur and Steve Graham. 2016. Writing research from a cognitive perspective

    2016

  22. [30]

    Masato Mita, Keisuke Sakaguchi, Masato Hagiwara, Tomoya Mizumoto, Jun Suzuki, and Kentaro Inui. 2022. Towards automated document revision: Grammatical error correction, fluency edits, and beyond. arXiv preprint arXiv:2205.11484

    2022 arXiv

  23. [31]

    Hiromi Narimatsu, Kohei Koyama, Kohji Dohsaka, Ryuichiro Higashinaka, Yasuhiro Minami, and Hirotoshi Taira. 2021. Task definition and integration for scientific-document writing support. In Proceedings of the Second Workshop on Scholarly Document Processing, pages 18--26

    2021

  24. [32]

    Robert C Nickerson, Upkar Varshney, and Jan Muntermann. 2013. A method for taxonomy development and its application in information systems. European Journal of Information Systems, 22(3):336--359

    2013

  25. [33]

    OpenAI. 2025. https://openai.com/gpt-5/ Gpt-5 is here . Accessed: 2025-10-3

    2025

  26. [34]

    James Pustejovsky, Harry Bunt, and Annie Zaenen. 2017. Designing annotation schemes: From theory to model. Handbook of Linguistic Annotation, pages 21--72

    2017

  27. [35]

    Nina Vandermeulen, Elke Van Steendam, Sven De Maeyer, and Gert Rijlaarsdam. 2023. Writing process feedback based on keystroke logging and comparison with exemplars: Effects on the quality and process of synthesis texts. Written Communication, 40(1):90--144

    2023

  28. [36]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, and 40 others. 2024. Qwen2 technical repo...

    2024 arXiv

  29. [37]

    Yi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang, Ruoxi Jia, and Weiyan Shi. 2024. https://doi.org/10.18653/v1/2024.acl-long.773 How johnny can persuade LLM s to jailbreak them: Rethinking persuasion to challenge AI safety by humanizing LLM s . In Proceedings of the 62nd Annual...

    2024 doi

  30. [38]

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

    2019 arXiv

  31. [39]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.