IBM Palette: Small Model Builds Custom Design Decks
IBM developed Palette, a system that fine-tunes a 20B open-weight model to generate slide decks in a specific design language.
Frontier models often produce slide decks with a uniform house style, while small open-weight models frequently generate broken layouts with vanishing titles or text running off the slide. IBM addressed this by developing Palette, a system designed to teach a small open-weight model to build decks in a specific design language while keeping ownership. The training data for the model follows the IBM Design Language. Small models struggle with slide layout because they must perform arithmetic to determine if text fits without seeing the rendered result. To solve this, IBM fine-tuned a 20B model using a method that splits the work between a designer role and a coder role. The coding stage is about 6–7 times faster because slides are coded in parallel after their context is fixed. The team used LoRA fine-tuning rather than full fine-tuning to teach the model a format and skill on top of a capable base.
The training data includes reasoning written as counterfactuals that weigh alternative choices and explain why they were rejected. Palette used a large open teacher model to generate around 1000 decks covering roughly 10,000 slides. Coder examples in the training data were about 5 times longer than designer examples, requiring designer examples to be oversampled 5 times. The system uses directives wrapped in [[ ]] to allow users to request specific design choices like changing a table to a timeline. The training data recipes include directives from 12 categories to help the model learn intent rather than fixed phrases.
Palette includes a harness that checks and repairs every slide before returning the deck to fix issues like overlapping text or malformed colors. The repair harness keeps a rewrite only if a checker confirms the slide improved, ensuring it cannot make a slide worse. Across 131 decks, the harness added an average of 3.5 points to the PresentBench score. Palette ranks first on Economics and Education domains on PresentBench, third on Talk, and second on the three-domain mean. Future plans for Palette include using checks as rewards for reinforcement learning to prevent mistakes in the first place.
Sources
Every sentence of this story was checked against facts quoted from these articles. How we make the news
Spotted an error? Tell us
