IBM trains a 20B open-weight model on about 1,000 synthetic decks: Palette tops PresentBench in Economics and Education
Synopsis
IBM researchers released Palette, which splits deck building into a designer that writes a structured brief and a coder that writes pptxgenjs code for each slide, both served by one LoRA adapter, and trains a 20B open-weight model on roughly 1,000 synthetic decks covering about 10,000 slides. The training data follows the IBM Design Language, and every generated slide is rendered and measured by a geometry checker, with repairs kept only when the checker confirms improvement. On PresentBench, Palette ranks first in Economics and Education, third in Talk, and second on the three-domain mean at 59.7 versus NotebookLM's 60.8, 1.1 points behind.
Interpretation
Palette lets a self-hostable 20B open-weight model produce decks in a specified design language rather than falling back to a frontier model's default style. Frontier models drift back to their own house style no matter how they are prompted, and confidential content may not be allowed to leave an organization's own systems. On PresentBench, Palette ranks first in Economics and Education, third in Talk, and second on the three-domain mean at 59.7, 1.1 points behind NotebookLM's 60.8 and ahead of the rest of the field.
The approach splits deck building into two roles: a designer writes a structured JSON brief, and a coder writes the pptxgenjs JavaScript that draws each slide from that slide's spec. A single pass has to make editorial judgments and geometric placements at once, which is why small models produce titles that vanish into the background, text running off the slide and stacked shapes. Both roles are served by one LoRA adapter; because each slide's context is fixed before any code is written, coder calls run in parallel, one per slide, making the coding stage about 6–7× faster.
Every synthesized deck is rendered and checked for overlaps and overflow by a geometry checker, and flagged slides go to a separate editor model that rewrites that slide's code. A rewrite is kept only if the checker confirms the slide improved, otherwise the original stays, so the harness is safe to run on every deck. Each PresentBench deck was generated once and scored with and without the harness; across 131 decks the harness adds 3.5 points on average.
Perspective
Evaluation covers the Economics, Education and Talk domains of PresentBench, with baseline numbers taken from the PresentBench paper; because PresentBench supplies a source document rather than a plan, in these runs a separate frontier model first drafted each plan from its document and Palette built the deck from that plan.
PresentBench measures whether a deck is good, not whether it follows a specific design language, and no known benchmark does the latter. On Talk, two closed systems score higher and Palette places third; a gap remains on figure-heavy decks.