PPTArena: a benchmark for PowerPoint editing

Editing presentations has historically been a tedious task, especially when aiming to maintain visual consistency across lengthy documents. With the rise of AI-based agents, new paradigms are emerging that promise to automate these complex tasks. In this context, the PPTArena benchmark represents a milestone by evaluating how agents modify real PowerPoint slides based on natural language instructions. Unlike previous approaches that focused on generating slides from scratch or interpreting PDFs, PPTArena proposes a set of 100 presentations with over 1,300 human-curated edits across 2,125 slides, covering text, graphics, animations, and professional master slide styles. Each edit is paired with a reference presentation and a target rubric, and evaluation is performed by two judges based on language and vision models (VLM): one measures instruction following based on structural differences, and the other assesses visual quality from slide images. Building on this, the PPTPilot agent demonstrates a structure-aware approach that plans semantic editing sequences, routes between programmatic tools and deterministic XML operations, and verifies each result in an iterative plan-edit-verify loop. This agent outperforms other models by more than ten percentage points on composite edits, with design changes and cross-slide modifications, achieving significant improvements in visual fidelity and overall consistency. However, all agents still face difficulties in long-range and document-scale tasks, highlighting how challenging reliable PowerPoint editing remains. This reality opens an opportunity for companies to adopt artificial intelligence solutions for businesses that integrate custom agents capable of automating complex document workflows. At Q2BSTUDIO, we develop custom applications that combine computer vision, natural language processing, and business logic to transform repetitive tasks into intelligent processes. Additionally, our expertise in AWS and Azure cloud services enables deploying these agents in a scalable and secure manner, while our capabilities in business intelligence (including Power BI) facilitate visualizing the performance of these agents. Cybersecurity is also key in these automated environments, and we offer audits and pentesting to protect corporate data. AI agents are the next step in process automation; at Q2BSTUDIO, we work to ensure companies can leverage their full potential without compromising quality or security.