ReviseMate: Interactive Review-Digestion Tool
- ReviseMate is an interactive system that digests and integrates reviewer feedback from STEM manuscripts, aiding revision plan formation.
- It employs a two-phase workflow with extraction, mapping, and categorization, supported by GPT-3.5 and BERT models for traceability.
- The system encourages human-AI collaboration by enabling user-driven organization of actionable comments to improve revision efficiency.
ReviseMate is an interactive system for digesting and integrating reviewer feedback on STEM papers during manuscript revision and rebuttal preparation. It was developed to address a review-digestion process characterized by time consumption, reading fatigue, and the need for comprehensive analytical skills, while avoiding the limitations of generic text-comprehension tools and prior review-analysis work that provide only high-level guidance (Xu et al., 21 Aug 2025). The system adopts a human-centered workflow grounded in Kintsch’s construction-integration theory and emphasizes human-AI collaboration, contextual linking between reviews and manuscript content, and user-controlled organization of extracted comments into revision plans (Xu et al., 21 Aug 2025).
1. Problem setting and design rationale
ReviseMate targets the phase in which authors assimilate reviewer reports and determine how those reports map onto manuscript changes. In the reported motivation, this phase is difficult because revision is iterative and often protracted, reviews may be unstructured, and accurate interpretation requires experience and strong analytical abilities (Xu et al., 21 Aug 2025). The system was designed after interviews and storyboard creation intended to identify authors’ concerns and methods for enhancing comprehension during review digestion (Xu et al., 21 Aug 2025).
The system’s design is explicitly human-centered and organized as a two-phase workflow. This workflow is described as being inspired by Kintsch’s construction-integration theory, which in this context motivates moving from extraction of actionable review content to contextual organization and planning (Xu et al., 21 Aug 2025). ReviseMate therefore does not treat review digestion as simple summarization. Instead, it decomposes the activity into preprocessing and analysis, with editable intermediate representations and repeated user verification.
A useful contextual distinction emerges when ReviseMate is compared with adjacent scientific-writing systems. ParaRev defines a paragraph-level scientific revision task in which a source paragraph and a revision instruction are used to generate a revised paragraph, and reports that detailed instructions improve automated revisions over general prompts (Jourdan et al., 9 Jan 2025). APRES, by contrast, uses LLMs to revise scientific papers against a discovered evaluation rubric and reports automated revision gains under citation-predictive criteria (Zhao et al., 3 Mar 2026). ReviseMate addresses a different locus in the workflow: not automated rewriting of manuscript text, but comprehension, decomposition, contextualization, and planning around reviewer feedback (Xu et al., 21 Aug 2025). This suggests that ReviseMate occupies an upstream cognitive-support role relative to systems that directly generate revisions.
2. Two-phase workflow and computational pipeline
ReviseMate is organized into a preprocessing phase and an analysis phase (Xu et al., 21 Aug 2025). The preprocessing phase is intended to extract actionable comments from raw review text, including reviews that are potentially unstructured. Three extraction methods are provided. In manual extraction, a user highlights and enters important review comments. In semi-automatic extraction, the system organizes reviewer-specific comment segments for user selection. In automatic extraction, GPT-3.5-turbo-16k summarizes and extracts key points into comment cards (Xu et al., 21 Aug 2025).
All extraction paths converge on a comment-card mechanism. Each extracted comment becomes an editable card that serves as an anchor for later analysis (Xu et al., 21 Aug 2025). To maintain traceability to the original reviews, ReviseMate applies BERT-based alignment between original and extracted comments, using BERT embeddings and cosine similarity to link concise reformulated cards back to the full review text (Xu et al., 21 Aug 2025). This link is central to the system’s claim of contextual support: compressed comments are not detached from their documentary provenance.
The analysis phase is designed to support organization, contextualization, and planning (Xu et al., 21 Aug 2025). First, each comment card can be assigned to hierarchical categories. These categories may be user-defined or drawn from defaults such as “Content,” “Workload,” and “Urgency,” and users can add, delete, or rename criteria and categories (Xu et al., 21 Aug 2025). Second, the system maps extracted comments to manuscript content. For each comment, ReviseMate uses bart-large-mnli to rank and suggest up to the top 5 most relevant paragraphs from the original paper (Xu et al., 21 Aug 2025). Users then review, confirm, or adjust these mappings, and can annotate thoughts or ideas directly at the mapped manuscript locations (Xu et al., 21 Aug 2025). Third, the system supports integrated organization for revision planning: comment cards and attached notes can be sorted or grouped by any criterion, and a Revision Editing sidebar supports assembling, rephrasing with LLM assistance, and exporting revision outlines or response drafts (Xu et al., 21 Aug 2025).
A notable constraint is that ReviseMate does not automatically generate revision plans (Xu et al., 21 Aug 2025). Users synthesize the final plan themselves from organized materials. This choice is consistent with the stated goal of lowering overreliance on AI by requiring human verification of major suggestions in comment extraction and paragraph mapping (Xu et al., 21 Aug 2025).
3. Interface architecture and interaction model
ReviseMate uses a split-view interface. The left panel displays the original paper, while the right panel manages review cards and the operations of extraction, categorization, mapping, and plan organization (Xu et al., 21 Aug 2025). The extraction interface lists comment cards and exposes multiple creation methods. Category customization is modal-based and augmented with visual cues and icons. Mapping suggestions are interactive: selecting a suggested paragraph scrolls to that paragraph in the manuscript, after which the user can confirm or revise the linkage (Xu et al., 21 Aug 2025).
The interaction model emphasizes manual verification rather than opaque automation. Drag-and-drop user-interface elements are used for mapping confirmation and adjustment (Xu et al., 21 Aug 2025). Annotation facilities allow users to move from identifying an issue in the reviews to developing candidate revisions or responses at the relevant manuscript location (Xu et al., 21 Aug 2025). The organization layer then supports grouping and sorting comment cards with notes, effectively constructing a revision-plan overview (Xu et al., 21 Aug 2025).
The reported field deployment led to adjustments in the interaction design. Users requested flexibility in how stages were traversed, and the deployed system consequently adopted freeform stage order rather than a forced sequence, supporting both individual and batch modes (Xu et al., 21 Aug 2025). In addition, the number of candidate mapping suggestions was reduced to alleviate choice overload observed in earlier studies (Xu et al., 21 Aug 2025). These changes indicate that the operational workflow was not assumed to be strictly linear even though the system is described as a two-phase design.
The interaction philosophy aligns with findings from adjacent revision-interface research. In ArgRewrite, a Wizard of Oz study of argumentative-writing revision interfaces found that detailed sentence-level revision categorization was the most helpful and the most effective condition, whereas finer sub-sentence feedback imposed higher complexity without clear added benefit (Afrin et al., 2021). ReviseMate likewise combines detailed process support with bounded presentation of candidates, especially in its paragraph-mapping step (Xu et al., 21 Aug 2025). A plausible implication is that review-digestion interfaces benefit from interpretable intermediate structures and constrained choice sets rather than maximal granularity.
4. Core support functions
ReviseMate’s support functions can be grouped into extraction, categorization, contextual mapping, and planning.
Extraction and traceability
Automatic extraction uses GPT-3.5-turbo-16k to summarize reviewer text into key points represented as comment cards (Xu et al., 21 Aug 2025). Semi-automatic extraction organizes reviewer-specific comment segments for user selection, while manual extraction allows direct highlighting and entry of important comments (Xu et al., 21 Aug 2025). BERT-base-uncased is used for semantic matching between original review passages and extracted comments, enabling users to refer back to the source material (Xu et al., 21 Aug 2025).
Customized categorization
Each comment card can be assigned to hierarchical categories that may be user-defined or drawn from defaults. Users can add, delete, and rename criteria or categories, and color coding is used to improve overview clarity (Xu et al., 21 Aug 2025). In the field deployment, categorization was reported to streamline comprehension and revision focus, and custom criteria were heavily used for domain-specific analysis (Xu et al., 21 Aug 2025).
Mapping to manuscript content
For contextual grounding, bart-large-mnli ranks paragraphs in the manuscript for relevance to each extracted comment, returning up to the top 5 candidates (Xu et al., 21 Aug 2025). Users inspect these candidates, confirm or revise the mapping, and can add annotations at the mapped locations (Xu et al., 21 Aug 2025). Trust in mapping suggestions was lower than trust in extracted comments, with a mean trust score in mapping of 3.56/7 in the controlled study, and this motivated limiting the number of presented candidates to five (Xu et al., 21 Aug 2025).
Organization and revision planning
ReviseMate supports sorting and grouping of comment cards by any criterion and provides a Revision Editing sidebar for assembling, rephrasing with LLM assistance, and exporting revision outlines or response drafts (Xu et al., 21 Aug 2025). However, the system does not automatically generate revision plans; ownership of the final synthesis remains with the author (Xu et al., 21 Aug 2025). In the field deployment, traceability and structured organization were reported to support collaborative revision and team communication (Xu et al., 21 Aug 2025).
These functions together define ReviseMate less as a summarizer than as a review-digestion workspace. This suggests an architecture centered on externalized intermediate representations—comment cards, categories, paragraph links, annotations, and grouped plan elements—rather than end-to-end generation.
5. Empirical evaluation
ReviseMate was evaluated in two stages: a controlled user study with and a field deployment with (Xu et al., 21 Aug 2025).
Controlled study
The controlled study used a between-subjects design comparing ReviseMate with a baseline consisting of side-by-side paper and raw review text without additional features (Xu et al., 21 Aug 2025). Participants analyzed actual reviews for a computer-science visualization paper and drafted a revision outline after a tutorial ensuring paper comprehension (Xu et al., 21 Aug 2025). Outcome measures included 7-point Likert ratings for usability, effectiveness, and trust or acceptance, and independent domain-expert ratings of outline completeness, accuracy, and organization on a 0–10 scale (Xu et al., 21 Aug 2025). Statistical analysis used the Mann-Whitney test because the data were non-parametric and the groups independent (Xu et al., 21 Aug 2025).
ReviseMate significantly outperformed the baseline on all reported effectiveness measures. The reported statistics were:
- Convenience of extraction:
- In-depth analysis access:
- Organization efficiency:
- Confidence in analysis:
- Outline satisfaction:
- Efficiency: (Xu et al., 21 Aug 2025)
Independent domain experts also rated the resulting revision outlines more favorably for the ReviseMate condition:
| Criterion | Control | ReviseMate |
|---|---|---|
| Completeness | 4.53 | 5.88 |
| Accuracy | 3.60 | 5.31 |
| Organization | 3.73 | 4.88 |
The reported -values were 0 for completeness, 1 for accuracy, and 2 for organization (Xu et al., 21 Aug 2025).
The study also documented interaction patterns. Participants alternated flexibly among extraction, categorization, and mapping, often combining steps and making context-driven adjustments rather than following a rigid sequence (Xu et al., 21 Aug 2025). Trust and acceptance were generally high for extracted comments and for the system overall, although mapping suggestions elicited more reservation (Xu et al., 21 Aug 2025). Qualitatively, novices relied more on defaults and AI suggestions, whereas experienced authors cross-checked and edited more vigorously (Xu et al., 21 Aug 2025).
Field deployment
The field deployment involved authors conducting real review-digestion tasks on papers whose reviews they had not seen for months (Xu et al., 21 Aug 2025). Users valued flexibility in extraction methods, reported that categorization streamlined comprehension and revision focus, and described AI-facilitated mapping as an “unstoppable trend” while still stressing the need for human oversight (Xu et al., 21 Aug 2025). Traceability and structured organization were reported to support collaborative revision and team communication, and participants explicitly requested multi-user collaborative editing and support for multiple review rounds in future releases (Xu et al., 21 Aug 2025).
Taken together, the reported results support the narrower claim that ReviseMate improved measured effectiveness for review digestion and revision-outline preparation relative to a raw-review baseline, rather than the broader claim that it automates manuscript revision itself (Xu et al., 21 Aug 2025).
6. Relation to revision research and system landscape
ReviseMate belongs to a broader family of revision-support systems, but its object of analysis is reviewer feedback rather than draft-to-draft edits. Several neighboring lines of work clarify this position.
Research on scientific revision corpora has shifted attention from sentence-level rewriting toward larger contextual units. CASIMIR provides 15,646 scientific articles with multiple author-integrated revisions, sentence-level alignments, paragraph metadata, and peer reviews, while also questioning the adequacy of standard revision metrics for this task (Jourdan et al., 2024). ParaRev moves explicitly to paragraph-level scientific revision and shows that detailed, paragraph-specific instructions yield significant improvements over general prompts across ROUGE-L, SARI, and BERTScore (Jourdan et al., 9 Jan 2025). EarlySciRev extends the resource base by extracting 578,440 validated early-stage paragraph-level revision pairs from commented-out LaTeX text in arXiv sources (Jourdan et al., 30 Mar 2026). These resources support modeling of manuscript revision, but they do not address the interactive comprehension of review reports that precedes or accompanies such revision.
Other systems focus on formative writing support or automated revision assessment. eRevise uses NLP features for rubric-based feedback on evidence use in student writing and improved evidence-related writing outcomes in a pilot deployment (Zhang et al., 2019). eRevise+RF extends this line by assessing evidence and reasoning revisions and providing revision feedback across three drafts, with deployment to 406 students (Liu et al., 1 Jan 2025). FeedbackWriter studies AI-mediated feedback for undergraduate essays and reports that TA-mediated adoption of AI suggestions improved revision quality in a randomized trial (Lu et al., 18 Feb 2026). These systems operate in educational contexts and focus on formative feedback and revision quality rather than manuscript-review digestion.
At the other end of the spectrum, APRES is an agentic LLM system that revises scientific papers against a discovered rubric predictive of future citations and reports that revised papers were preferred over originals in 79% of human expert comparisons (Zhao et al., 3 Mar 2026). ReviseMate does not attempt this form of automated paper rewriting. Instead, it provides infrastructure for extracting, organizing, and contextualizing review comments so that authors can produce their own revision plans (Xu et al., 21 Aug 2025). This distinction is central: ReviseMate augments deliberation around reviewer feedback, whereas APRES automates manuscript-level revision under an explicit optimization target.
Visualization research also offers a useful parallel. The Cumulative Revision Map represents version-controlled document evolution as a sparse matrix over revision number and document position, revealing authoring and editing patterns (Kim et al., 2012). ReviseMate is not a revision-history visualizer, but its mapping and organization functions similarly externalize structure so that users can inspect where comments apply and how revision work is distributed (Xu et al., 21 Aug 2025). This suggests a shared emphasis on interpretability of revision processes rather than opaque end-state generation.
7. Limitations, boundaries, and future directions
Several limitations are explicit. First, ReviseMate’s mapping assistance required trust calibration: mean trust in mapping was 3.56/7, lower than trust in extracted comments, which led to limiting suggestions to five candidate paragraphs and preserving manual verification (Xu et al., 21 Aug 2025). Second, the system is intentionally not fully automated. No automatic revision plan is generated, and human confirmation is required for major system suggestions (Xu et al., 21 Aug 2025). Third, field users requested support for multi-user collaborative editing and multiple review rounds, indicating that the reported version did not yet fully support common coauthor workflows (Xu et al., 21 Aug 2025).
The broader scientific-revision literature indicates additional challenges. CASIMIR shows that common automatic metrics such as BLEU, SARI, ROUGE-L, and even BertScore can misalign with the pluralistic nature of revision, with CopyInput often rivaling or exceeding learned models on standard metrics (Jourdan et al., 2024). ParaRev similarly notes that current evaluation metrics can penalize valid alternative revisions and emphasizes the need for more robust reference-less or multi-reference evaluation (Jourdan et al., 9 Jan 2025). These results do not directly evaluate ReviseMate, but they suggest that downstream assessment of review-digestion tools may require richer measures than overlap-based proxies if such systems are later connected to manuscript revision generation.
Future directions reported for ReviseMate include multi-user collaboration and support for multiple review rounds (Xu et al., 21 Aug 2025). A plausible implication is that coupling ReviseMate’s review-digestion layer with resources such as CASIMIR, ParaRev, or EarlySciRev could support end-to-end systems spanning review comprehension, revision instruction formation, and manuscript rewriting (Jourdan et al., 2024, Jourdan et al., 9 Jan 2025, Jourdan et al., 30 Mar 2026). The current system, however, is defined by a narrower and more interpretable objective: helping authors digest reviews by extracting comments, contextualizing them within the manuscript, organizing them into user-defined structures, and supporting manual synthesis of revision plans (Xu et al., 21 Aug 2025).
