The Pulse
IBM Research Trains Palette, a 20B Model for Slide Decks
IBM Research introduced Palette, a fine-tuned 20B model for building slide decks in the IBM Design Language. On PresentBench, the team reports first-place rankings in Economics and Education and a 3.5-point average gain from its slide-repai
AI.info Team · Written with AI · fact-checked ·
Across 131 PresentBench decks, IBM Research says its check-and-repair harness added 3.5 points on average. Palette does not send its first draft straight to users: it renders slides, checks for layout faults and sends flagged code to a separate editor model. A rewrite stays only if the checker confirms the slide improved.
Palette divides planning from slide geometry
IBM Research published its account of Palette on November 3, 2026. The team describes a 20B model trained to follow the IBM Design Language, with one role planning the presentation and another drawing each slide.
“We fine-tuned a 20B model to build decks the way people do: a designer plans the deck and a coder draws each slide.”
Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo, IBM Research
The designer turns a plan into a structured JSON brief specifying the palette, type system, slide layouts, content and design rationale. The coder takes each slide’s specification and writes the JavaScript that draws it. Because the slide context is set before coding begins, those coding calls can run in parallel; IBM says that makes the coding stage about 6–7 times faster.
The team trained a LoRA adapter rather than applying full fine-tuning. Coder examples ran about five times longer than designer examples, so the team oversampled designer examples fivefold to avoid undertraining that role.
About 1,000 synthetic decks supplied the training material
IBM Research says it generated around 1,000 decks covering roughly 10,000 slides because it lacked a public corpus containing deck plans, design decisions and slide code. Each example includes a counterfactual reasoning trace: it weighs an alternative and explains why the design chose another option.
The team rendered each synthetic deck and used its checker to find overlaps and overflow. It repaired broken slides, then rewrote the examples’ reasoning to describe what failed and why the correction worked. IBM identifies factual faithfulness as the hardest problem: prompt rules alone did not prevent a model from adding facts absent from its input.
In the team’s tests, reference decks changed the teacher model’s output more reliably than extra prompt rules. That finding shaped the data process: the researchers fixed reference examples when the generated output was wrong.
Palette ranks first in two PresentBench domains
PresentBench scores generated decks on fundamentals, visual design, completeness, correctness and fidelity. For this evaluation, a separate frontier model converted each benchmark source document into a plan before Palette built the deck; the comparison uses baseline scores from the PresentBench paper.
Palette ranked first in Economics and Education, third in Talk, and second across the three-domain mean. Its average was 59.7, compared with 60.8 for NotebookLM, a 1.1-point gap. IBM says two closed systems scored higher than Palette in Talk, where the decks are narrative and speech-driven.
The benchmark measures deck quality, but not whether a system follows a particular design language. IBM says it is not aware of a benchmark for that specific capability.