IBM Research fine-tunes a 20B model to build slide decks in the IBM Design Language

IBM Research has published Palette, a 20B-parameter model fine-tuned to build slide decks in a specific design language. According to IBM Research, Palette ranks first on two of the three domains it measured on PresentBench. Its training data follows the IBM Design Language.

The work was published on the Hugging Face blog on November 3, 2026, by Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo.

The premise

IBM Research argues that frontier models drift back to their own house style when asked for decks, regardless of prompting, and that some content is too confidential to leave a company's systems. Its proposed alternative is a small open-weight model, trained on synthetic data in your own design language and wrapped in a checking harness. IBM says small open models fail at deck building in predictable ways: titles that vanish into the background, text running off the slide, and overlapping shapes.

IBM attributes this to three difficulties. The model writes code that places each element at an x, y, width and font size without seeing the result. It must make editorial and geometric decisions in one pass across 10 to 20 slides. And no public data shows the rationale or code behind finished decks.

Designer and coder roles

Palette splits the work into two roles served by a single LoRA adapter:

• Designer: reads the plan and writes a structured JSON brief containing the palette, the type system, and each slide's layout, exact content and a short design rationale.
• Coder: takes one slide's spec from the brief and writes the pptxgenjs JavaScript that draws it, preceded by its reasoning.

Because each slide's context is fixed before code is written, coder calls run in parallel, one per slide. IBM says this makes the coding stage about 6 to 7 times faster. The brief is readable and editable, so a person or agent can change one slide's layout without regenerating the deck.

IBM chose LoRA over full fine-tuning because it teaches a format and skill without disturbing the base model's knowledge, and it made each experiment cheaper. Coder examples are about 5 times longer than designer examples, so a naive mix undertrained the designer until IBM oversampled designer examples 5 times.

Synthetic training data

IBM generated around 1,000 decks covering roughly 10,000 slides with a large open teacher model. Each deck starts from a recipe that sets genre, length, layouts and any user directives. Every example carries counterfactual reasoning that weighs an alternative and explains why it was rejected.

Every synthesized deck is rendered and checked for overlaps and overflow by the same checker the harness uses at inference. Broken slides are repaired, and their reasoning is rewritten to explain what the original layout got wrong.

IBM reports that faithfulness, meaning no facts absent from the input, was the hardest problem, and that prompt rules did not prevent invented content. Reference decks in the teacher's prompt did: fixing a reference reliably changed the output, while adding a rule rarely did.

Directives

Users can steer a deck with short instructions wrapped in double square brackets inside the plan, such as [[make this a timeline]]. The training recipes include directives from 12 categories, each phrased several ways. The categories shown include layout, tone ([[be candid, no spin]]), audience ([[this is for the board]]), order ([[lead with the ask]]), leaving out ([[no pricing]]) and emphasis ([[highlight the headline number]]).

Check-and-repair harness

Palette does not return raw model output. Simple rules first fix known mistakes, such as a malformed color code that would turn every slide black. Each slide is then rendered and measured by a geometry checker that flags text that overlaps or gets clipped.

Context window, pricing and release status of the weights are not stated in the post. The PresentBench ranking is IBM Research's own claim.

What this means

The design separates editorial decisions from geometry, which gives each role one kind of decision and leaves an editable JSON brief between them. IBM's reported results on synthetic data point to two practical levers: reference decks steered the teacher model more reliably than written rules, and using the same checker in data creation and at inference turned each caught mistake into a training example. The first-place PresentBench result on two of three domains is IBM's claim, and the post presents Palette as a test of whether a self-hosted model can compete with frontier systems while the owner keeps control of the result.