EditPPT: Faithful Long-Deck Slide Editing via Structured Tool-Using Multi-Agent with Dual-Modal Validators
EditPPT is a multi-agent framework that achieves faithful long-deck slide editing by reformulating the task as a constrained tool-selection problem using native PowerPoint operations and dual-modal validation, demonstrating superior accuracy and robustness on the newly introduced DeckEdit-Bench.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern world, the slide deck is a universal language of business, science, and education. It is the medium through which complex ideas are organized, presented, and often sold. For decades, creating these presentations has been a manual, labor-intensive task, requiring a human to click, drag, and type every element into a digital file. Recently, artificial intelligence has begun to automate the creation of these decks from scratch, turning text descriptions into visual slides. However, a different and perhaps more difficult challenge has remained unsolved: editing an existing deck. When a user asks an AI to change a specific chart on page twenty or rephrase a title on page five, the system must not only find that exact spot but also leave every other part of the presentation untouched. This requires a delicate balance of precision and restraint, ensuring that a small request does not accidentally ruin the layout of the entire document.
A team of researchers at KAIST AI has developed a new system called EDITPPT to solve this problem of faithful editing. Rather than asking an artificial intelligence to write code that might accidentally break the file, the researchers designed a system that treats the presentation like a physical object with distinct, manageable parts. The system breaks the user's request down into a series of specific, small tasks. It then assigns these tasks to specialized digital workers, each trained to handle only one type of object, such as a text box, a table, a chart, or a shape. These workers do not guess or invent; they use a direct, reliable connection to the PowerPoint software to make the exact change requested. Before the system considers the job done, it checks its work twice: once by reading the digital structure of the file to ensure the right words were changed, and again by looking at a picture of the slide to make sure the visual layout still looks correct.
The researchers tested this approach on a new collection of real-world presentations, including decks with more than thirty slides, which are often where automated systems fail. They found that their method successfully completed nearly every editing task it was given, with a success rate of 99.5 percent. More importantly, the system was remarkably good at preserving the parts of the deck that were not supposed to change. It correctly identified the target slides 88.7 percent of the time and followed the user's instructions 82.5 percent of the time, all while keeping 91.5 percent of the unrelated objects and formatting exactly as they were. In contrast, other systems that rely on writing code or manipulating raw file data often struggle as the presentation gets longer, making unintended changes to slides that should have been left alone.
The core of this success lies in how the system interacts with the presentation file. Instead of trying to understand the entire file at once, which can be overwhelming and prone to error, the system focuses on one slide at a time. It translates the visual elements of the slide into a clear, structured list that the AI can understand. When a user asks to change the color of a specific box, the system identifies that box by its unique ID and sends a command to a specialized agent that knows exactly how to change colors without touching the text next to it. This approach prevents the kind of cascading errors that happen when a system tries to rewrite a large section of code and accidentally breaks something else nearby. The researchers also introduced a new benchmark, a set of twenty-eight real decks with hundreds of slides and hundreds of editing requests, to prove that their system works on the messy, complex files people actually use, not just on perfect, computer-generated examples.
The results show that this method is particularly strong when dealing with long, complex presentations. While other systems saw their performance drop significantly as the number of slides increased, EDITPPT maintained its high level of accuracy and preservation. This suggests that the key to reliable editing is not just having a smarter AI, but having a smarter way of using the AI. By limiting the AI's actions to a specific set of tools and having it check its work through both digital and visual lenses, the system achieves a level of reliability that was previously missing. The researchers conclude that for AI to truly become a useful partner in document editing, it must be guided by strict constraints that match the structure of the application it is using, rather than being given the freedom to rewrite the file however it sees fit. This approach ensures that the final result is not just a changed file, but a faithful version of the original that respects the user's intent and the integrity of their work.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.