IBM Research launched Palette, a system designed to create slide decks that strictly adhere to a specific design language. The tool splits the slide creation process into two distinct roles handled by a single LoRA adapter, assigning one part to editorial judgment and another to geometric layout. This approach allows the model to manage both the content style and the structural arrangement of the presentation.
The system fine-tunes a 20B open-weight model using synthetic data and a custom harness to maintain output quality. The harness checks generated slides for overlaps and overflow, rewrites code for broken slides, and ensures facts trace back to the brief. Researchers generated around 1,000 decks covering 10,000 slides to train the model on design decisions and layout constraints. On the PresentBench benchmark, Palette ranked first on Economics and Education domains and second overall against NotebookLM and Manus 1.6.
Developers and design teams focused on automated presentation workflows should pay attention to this release. The ability to enforce strict design languages and catch layout errors before a slide is finalized could save significant time in professional settings. This tool offers a new way to standardize presentation output while maintaining high editorial and geometric standards.
