IBM's Palette splits decks into a designer and a coder, and a 20B model keeps up

IBM Research has published Palette, a 20B open-weight model fine-tuned to build slide decks in a specific design language, in a Hugging Face blog post. The premise is concrete: ask a frontier model for ten decks and they converge on the same tasteful house style, and some content is confidential enough that it cannot leave your own systems at all.

Decks are hard for small models for a reason. The model does not draw a slide; it writes code that places every element at a given x, y, width and font size. Whether a 75-character title fits on two lines at 56pt is arithmetic it has to get right without seeing the result, and it only finds out it was wrong after the code renders.

Palette splits the job the way a person would. A single LoRA adapter serves two roles: a designer reads the plan and writes a structured JSON brief — palette, type system, and for each slide its layout, exact content and a one-line rationale — and a coder turns one slide's spec into pptxgenjs JavaScript, with its reasoning written first. The designer handles editorial judgment, the coder handles geometry.

The split also pays off in engineering terms. Because each slide's context is fixed before any code is written, coder calls can run in parallel, making the coding stage roughly six to seven times faster. IBM also hit a training snag: coder examples run about five times longer than designer examples, so a naive mix under-trained the designer until its data was oversampled fivefold.

The training data is synthetic. The team used a large open teacher model to generate around 1,000 decks covering roughly 10,000 slides, each carrying counterfactual reasoning that records not just the choice but the alternative that was rejected and why. Every generated deck is rendered and checked for overlaps and overflow by the same checker used at inference; broken slides are repaired and their reasoning rewritten, so each caught mistake becomes a better training example.

At inference there is a check-and-repair harness. Rules fix known mistakes first, such as a malformed color code that would turn every slide black; then each slide is rendered and measured, and flagged slides go to a separate editor model that rewrites the code. A rewrite is kept only when the checker confirms the slide improved, which makes the harness safe to run on every deck.

On 131 PresentBench decks, the harness adds 3.5 points on average. Scored across three domains, Palette ranks first on Economics and Education, third on the more narrative Talk domain, and second overall at 59.7 versus NotebookLM's 60.8 — ahead of the rest of the field as a single 20B open-weight model.

IBM says the next step is to turn the harness's cheap, verifiable signals into reinforcement-learning rewards so the model stops making these mistakes in the first place; because the coder works one slide at a time, each slide is a short, independently checkable rollout suited to GRPO-style training. It is also exploring figure extraction with Docling and a lighter path that re-skins output from one or two existing .pptx files without retraining.