IBM taught a 20B model to build decks. The verification loop does the decisive work

IBM Research has introduced Palette, a 20B open-weight model fine-tuned with LoRA to produce decks in the IBM Design Language. One adapter serves two roles. A designer writes a structured JSON brief, then a coder produces PptxGenJS JavaScript for each slide.

The training set was synthetic. The team generated about 1,000 decks containing roughly 10,000 slides and added explanations of rejected design alternatives. Coder calls run in parallel for individual slides, making that stage about 6 to 7 times faster according to the authors.

Generation does not stop at the model's first attempt. A harness renders the deck and detects overlaps, clipped text and elements outside the canvas. An editor rewrites a faulty slide, but the system keeps the revision only when the checker measures improvement. Across 131 decks, this process added 3.5 points on average.

Palette turns one difficult task into an editable brief and slide code

The model no longer has to decide the story, exact content and geometry of a whole deck at once. The designer sets typography, colors, layout and content for every slide. The coder then handles each page separately.

The intermediate result is readable by a person or another agent. A team can change one argument or one layout without repeating the entire generation. That boundary is more transferable than the presentation format itself, because the same approach can support charts, diagrams or user interfaces.

Companies gain more control from a checker than another round of prompt engineering

IBM found faithfulness to be the hardest problem. If a brief requested a 6 by 5 table but supplied only three bullets, the coder invented the missing content. The team reports that adding another prompt rule helped less than fixing the reference decks in the training examples.

That is the practical enterprise lesson. Quality depends on more than choosing a larger model. It also comes from a structured intermediate format, verifiable rules and a repair process capable of rejecting its own proposal. Open weights additionally allow confidential documents to remain on company infrastructure.

The benchmark rewards good decks but does not test the main design language claim

Palette scored 59.7 on PresentBench. It ranked second on the mean of three reported domains, 1.1 points behind NotebookLM at 60.8. It led Economics and Education but placed third in Talk.

The comparison has two important limits. PresentBench does not measure fidelity to a particular design language, even though that is Palette's central goal. A separate frontier model also converted each source document into a plan before Palette built the deck. The result therefore does not describe a fully local end-to-end system powered only by the 20B model.

Company templates, factual fidelity and figure-heavy slides are the real tests

IBM plans to use checker signals as reinforcement learning rewards, add figures extracted from source documents and explore reskinning from one or two existing .pptx templates without retraining. That last step will directly test adaptation to someone else's design language.

A production deployment needs checks beyond geometry. It must trace facts to sources, enforce approved fonts and colors, handle logos and accessibility, and preserve content during repair. Only then does a research pipeline become a system trusted with a client presentation.

Lilith's verdict

Palette wins by sending an inspector with a ruler after every slide and allowing that inspector to reject a repair. That small authority separates an impressive demo from usable production.