PaperDebugger: A Plugin-Based Multi-Agent System for In-Editor Academic Writing, Review, and Editing

1 Introduction

• External tools can fragment writing flow and lose revision provenance, while in-editor suggestions such as Writefull offer limited transparency into applied changes.Copy-and-paste workflows separate feedback, applied changes, and reasoning from the document.Source: §1 · p. 1
• PaperDebugger integrates critique, refinement, and revision into Overleaf, tying persistent interaction history, patch-based edits, and structure-aware feedback to document context.The system is implemented as a Chrome extension communicating with a Kubernetes-based backend, with MCP enabling modular functionality.Source: §1 · p. 1
• Deployment to real users and anonymized telemetry indicate sustained engagement and active adoption in live academic writing scenarios.The demonstrated workflow includes requesting critiques on selected passages, reviewing diff-style revisions, and applying accepted patches.Source: §1 · p. 2
• The system combines in-editor assistance, Kubernetes-orchestrated multi-agent execution, MCP-powered retrieval, structured review, and deterministic diff-based editing.The stated contributions also include deployment through the Chrome Web Store and early telemetry of repeated critique and revision use in authentic writing settings.Source: §1 · p. 2

2.1 Architecture Overview

• The five-layer architecture comprises presentation, backend, agent, protocol, and infrastructure layers, separating Overleaf interaction from workflow execution and operational services.The backend coordinates authentication, session state, and routing; the protocol handles structured communication, while infrastructure provides storage and operational services.Source: §2.1 · p. 2, Figure 1 · p. 2
• The Chrome extension captures selected text and project state, presents edits as before–after diffs, and applies accepted patches directly to the LaTeX source.Its floating panel and inline controls integrate into Overleaf, avoiding copy–paste cycles and maintaining revision history.Source: §2.1 · p. 2
• The Go backend runs stateless LLM agents in isolated Kubernetes pods for high concurrency and horizontal scaling, while the streaming protocol delivers intermediate outputs in real time.The protocol is compatible with OpenAI’s server-sent event format; orchestration also handles routing, model selection, permissions, and schema validation.Source: §2.1 · p. 2
• PaperDebugger supports prompt-template agents for low-latency tasks and workflow-based agents that coordinate LLM calls, tools, and validation for complex tasks.Examples include grammar polishing for template agents and deep research, paper retrieval, and full-document enhancement for workflows.Source: §2.1 · p. 2

2.2 Agentic Design

• XtraMCP supplies validated academic-writing tools, including semantic literature retrieval, segment-level AI review, and XtraGPT, a model suite tuned for scholarly revisions.Internal schemas and consistency checks are intended to minimize hallucinations.Source: §2.2 · p. 2
• Specialized Reviewer, Enhancer, Scoring, and Researcher agents handle critique, refinement, evaluation, and literature lookup; full-document review is split across worker pods and merged.The coordinating agent decomposes a review request into segment-level sub-queries before dispatching them.Source: §2.2 · pp. 2–3

Usage Analytics

• Telemetry from May 2025–January 2026 was anonymized and collected from both the Chrome extension and backend.The analysis examines usage in real writing environments.Source: §3 · p. 3

Real-World Adoption

• 4,116 extension installs, 2,761 registered users, and 732 monthly active users accompanied 3,106 projects and 7,447 writing threads.The authors interpret project and thread counts as sustained multi-session usage; user feedback also noted domain-specific tone and performance drops on long documents.Source: §3 · p. 3, Table 1 · p. 3

Interaction Patterns

• Users repeatedly refined within sessions, frequently inspected patch diffs before applying changes, and showed sustained in-editor revision activity.The paper describes revision rather than one-shot generation as the dominant interaction pattern.Source: §3 · p. 3, Table 2 · p. 3

4.1 In-Editor Editing and Patch

• Authors select text in Overleaf, request critique or refinement, then inspect and apply before-after patches generated by coordinated agents.Inline previews let authors review suggestion rationales and choose a revision.Source: §4.1 · p. 3, Figure 3 · p. 4
• The editing case demonstrates a transparent patch workflow intended to improve clarity and structure while retaining an efficient in-editor experience.Source: §4.1 · p. 3

4.2 Deep Research and Comparative Analysis

• Deep research searches arXiv and curated corpora for ranked papers with metadata, abstracts, and generated relevance explanations, then compares a selected paper with the author’s work.The comparison covers goals, datasets, methods, evaluation protocols, and limitations, and can produce a citation-ready summary table.Source: §4.2 · pp. 3–4
• Broader searches consolidate related works into research maps with topical clusters and takeaways for positioning the manuscript.Source: §4.2 · p. 4
• Deep-research outputs, including literature retrieval and comparison materials, are delivered in-editor without manual copying.The workflow also supports research-level reasoning and literature synthesis within the writing environment.Source: §4.2 · p. 4

• PaperDebugger integrates critique, structured review, literature retrieval, and diff-based editing into an Overleaf writing environment.Source: §5 · p. 4, Figure 4 · p. 4
• Real-user deployment and early usage analytics indicate sustained engagement, while case studies cover micro-level refinement and deep-research tasks.Source: §5 · p. 4
• The manuscript states that PaperDebugger polished portions of the text, while no scientific content was generated by AI.Source: §5 · p. 4