Researchers have released PPTArena, a benchmark designed to measure how well AI agents handle real PowerPoint editing tasks.
The benchmark includes 100 slide decks with over 1,300 human-curated edits spanning text, charts, animations, and professional master styles across 2,125 slides. Each edit is scored by two Vision-Language Model judges — one checks whether the agent followed the instruction, the other rates visual quality. Alongside the benchmark, the team released PPTPilot, an agent that plans edit sequences and routes between programmatic tools and XML operations, checking its own work in a loop. PPTPilot outperformed competing VLM-based agents by more than 10 percentage points on compound, layout-sensitive, and cross-slide edits.
Most AI document-editing benchmarks lean on image or PDF renderings, or test text-to-slide generation from scratch — neither reflects what office workers actually do when they open a deck and start changing things. PPTArena's focus on editing existing slides is a more honest test of practical utility, and the 10-point gap between PPTPilot and general VLM agents suggests that structure-aware planning matters more than raw model capability here.
The catch: even PPTPilot stumbles on long, document-scale tasks, which is exactly the kind of editing that would save someone an afternoon — so the benchmark's main finding is that the problem remains largely unsolved.