SlideLab uses AI agents to build and test research presentations

Making slides from a research paper involves more than shrinking its text. A presentation needs a clear story, useful visuals and an order that helps listeners understand the work. SlideLab, a system described in an arXiv paper published September 28, 2026, is designed to handle those tasks.

The system uses several AI agents, each assigned part of the work. It first plans the presentation’s narrative, then creates and revises a shared slide deck. The authors describe the approach as “training-free,” meaning it does not train a new AI model for this task.

How does it make slides?

SlideLab’s agents handle content planning, visual creation, layout refinement and checking whether claims are supported by the source paper. They work in repeated rounds, rather than producing a deck in one pass. The goal is to make a coherent presentation while keeping its content grounded in the research.

In a blind human preference study, the authors compared SlideLab with both open-source and commercial systems. They report that participants preferred SlideLab’s presentations on 77% of the papers tested. The system also used roughly four times fewer inference tokens than the strongest open-source comparison system. Tokens are small pieces of text processed by an AI model, and the figure is a measure of model usage—not a direct measure of cost.

How are presentations evaluated?

The paper also introduces ConfArena, an evaluation method that simulates a conference audience and assesses a presentation one slide at a time. The authors say it produces system rankings that match human rankings and can spot deliberate problems, such as altered numbers, degraded figures, missing slides or slides placed in the wrong order.

The paper reports promising results, but the abstract does not give details such as the number of papers or people in the preference study. Those details matter when judging how broadly the findings apply. ConfArena is presented as a way to test slide decks, not as evidence that an AI-generated deck is ready to present without a human review.