A new AI framework claims it can turn a dense research paper into a conference-ready slide deck almost as well as a human - using a fraction of the compute of its closest rival.
Researchers built SlideLab, a multi-agent system that plans a presentation's narrative, drafts slides, generates visuals, and checks its own work against the source paper before finalizing a deck. Rather than one model doing everything, separate agents handle content planning, layout, and fact-checking, then iterate on a shared draft. In a blind human preference study, evaluators picked SlideLab's slides over both open-source and commercial alternatives for 77% of the papers tested. The team also built ConfArena, a companion evaluation tool that simulates a conference audience grading each slide and can catch planted errors like falsified numbers or shuffled slide order.
Turning papers into slides is grunt work every researcher dreads, and the pitch here is speed without sloppiness: SlideLab reportedly used about four times fewer inference tokens than the strongest open-source alternative it beat. That efficiency claim matters more than the preference score - cheaper generation is what would actually get a tool like this used at scale, rather than staying a one-off demo.
The 77% figure comes from the paper's own study, not an independent bake-off, so it's a promising early result rather than a verdict. The real test will be how SlideLab holds up once it's competing with the AI slide-deck tools already crowding academic conferences.