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

What This Paper Is About

Current AI writing assistants, powered by Large Language Models (LLMs), often feel disconnected from the actual writing process. For academics using specialized editors like Overleaf, this means constantly copying text, pasting it into a separate AI tool like ChatGPT, and then bringing the suggestions back into their document. This interrupts the flow of writing and makes it hard to track changes.

This paper asks the question: How can we build an AI writing assistant that works directly inside the editor, feeling like a natural part of the writing workflow? The problem matters because a disconnected tool is inefficient and breaks a writer's concentration. By bringing the assistant into the editor, the goal is to create a seamless, powerful, and context-aware writing partner that understands the document's structure and history.

Key Approach

The researchers built a system called PaperDebugger, which works as a browser extension for Chrome. It integrates directly into the Overleaf editor, a popular tool for writing academic papers. This eliminates the need for copy-pasting and allows writers to use AI assistance without ever leaving their document.

The approach is built on a multi-agent system. Think of this as having a team of specialized AI assistants. One agent is an expert in reviewing text for clarity, another specializes in rephrasing sentences, and a third can search for relevant research papers. When a user highlights a piece of text, they can choose the right agent for the job. These agents are managed by a powerful backend system that can handle complex requests, like deep literature reviews, and run multiple tasks at once.

What makes this approach novel is its deep integration. PaperDebugger can read selected text, send it to the appropriate AI agent, and then display the proposed changes as a clean "before-and-after" comparison. The writer can review the suggestion and apply it with a single click.

Main Findings

The research team successfully developed PaperDebugger and made it publicly available on the Chrome Web Store. By analyzing anonymous usage data, they found that it is not just a theoretical concept but a practical tool that people actively use. With over 100 installations and 78 registered users, about a third of them used the tool regularly, indicating it provides real value.

The data revealed key user behaviors. Writers didn't just ask the AI to generate text from scratch. Instead, they used it to refine and edit existing text. The most common actions were:

• Viewing proposed changes (diffs): Users viewed over 1,000 "before-and-after" comparisons.
• Applying patches: Users applied a suggested change directly into their document over 350 times.

These findings show that writers want control and transparency. They highly value the ability to see exactly what the AI is suggesting before accepting a change. The frequent, repeated use of the tool within writing sessions proves that an in-editor assistant can successfully become part of a writer's regular workflow.

Why This Matters

This paper offers a practical blueprint for the future of AI-powered writing tools. By moving away from external chatbots and embedding assistance directly into the writing environment, PaperDebugger shows how AI can become a true collaborator rather than just a simple grammar checker. This seamless integration can significantly improve a writer's efficiency and focus.

The implications go beyond academic writing. The principles demonstrated here—in-context assistance, specialized agents, and user-controlled revisions—could be applied to other professional domains like legal document drafting, software development, and technical writing. This work points toward a new generation of smart tools that enhance human expertise instead of trying to replace it, making complex tasks faster and more manageable.