IBM’s Palette tunes a 20B model on about 1,000 synthetic decks and 10,000 slides, separating design JSON from slide code. It checks rendered output for overlap and overflow and retains repairs only when they help. Its three-domain PresentBench average was 59.7, placing second, but a separate frontier model drafted plans from input documents and baseline scores came from the earlier paper. The evaluation does not measure adherence to a specific corporate design language, so the result illustrates a structured generation-and-checking approach rather than validating that capability.
Back to Tech
IBM Palette separates slide design and coding with a 20B model
A separate frontier model supplied evaluation plans, and the benchmark does not measure adherence to a company's design language.