New AI Model Palette Takes First Place in Two Categories of Presentation Test
The team described Palette—a model with open weights and 20 billion parameters for creating presentations in the IBM Design Language style. In PresentBench, it secured first place in two out of three domains.
The developers fine-tuned Palette using synthetic data created in accordance with the IBM Design Language. The model divides the presentation creation process into two stages: first, it generates a structured plan with slide content and design, then writes JavaScript code to assemble them via pptxgenjs.
Before delivering the final output, a special module checks the slides for overlaps, cropped text, and elements extending beyond boundaries. A separate model rewrites the code for problematic slides, but the correction is retained only after confirmation that the result has improved.
In testing on 131 presentations, this module, on average, improved the score by 3.5 points. According to the authors, Palette took first place in the Economics and Education categories, third in Talk, and second in the average score across the three PresentBench domains.
