Palette: teaching a small model to build decks in your design language

IBM Research has introduced Palette, a fine-tuned 20B open-weight model that generates slide decks in an organization's own design language, according to a November 3 post on Hugging Face by Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo. The authors argue that frontier models drift back to a recognizable house style and that confidential material sometimes cannot leave company systems, while small self-hosted models tend to produce broken layouts.

Palette splits the job into two roles served by one LoRA adapter. A designer writes an editable JSON brief covering palette, type and each slide's content, and a coder writes pptxgenjs code per slide, in parallel, making that stage about 6 to 7 times faster. Training data came from a large open teacher model that produced around 1,000 decks and roughly 10,000 slides following the IBM Design Language, each annotated with reasoning about rejected alternatives. Users can steer output with bracketed instructions in the plan, drawn from 12 categories.

Every deck also passes through a harness that renders slides, flags overlaps or overflow, and keeps a separate editor model's rewrite only if a checker confirms improvement. IBM says this added 3.5 points on average across 131 decks.

On PresentBench, the team reports first place in Economics and Education, third in Talk, and second overall on the three-domain mean, at 59.7 versus NotebookLM's 60.8. Baselines came from the benchmark's paper, and a separate frontier model drafted Palette's plans. The authors note that no benchmark they know of measures adherence to a design language. They plan to reuse the checks as reinforcement learning rewards.