Palette: teaching a small model to build decks in your design language

Slide decks sound trivial until a company wants every one to look like its own template, or the content is too sensitive to send to an outside model. Frontier systems drift back to their house style. Small self-hosted models usually produce broken slides: vanishing titles, overflowing text, stacked shapes.

Palette, from IBM Research, fine-tunes a 20B model with a LoRA adapter on synthetic decks that follow the IBM Design Language. On PresentBench it ranks first on the Economics and Education domains, third on Talk, and second overall, 59.7 against NotebookLM's 60.8.

If that holds up, owning a deck model that speaks your brand becomes realistic for organizations that cannot or will not ship content to a frontier API.

We like this work because it is built the way good teams actually work. A designer role writes a structured brief: palette, type, layout and a short rationale per slide. A coder role writes each slide's code, in parallel, one call per slide. The brief is readable and editable, so a person can change one slide without regenerating the deck. That is useful AI: smaller, inspectable, controllable.

The harness is the part we'd steal. Every slide is rendered and measured for overlaps, clipping and overflow. Flagged slides go to an editor model, and a rewrite is kept only if the checker confirms it improved. Across 131 decks, the harness adds 3.5 points on average. A fix that can't make things worse is a fix you can run on everything.

The team is also candid about the hardest failure: slides containing facts that were never in the input. They found prompt rules barely moved the teacher model; fixing the reference decks did. That is a sharp, transferable lesson, and also a warning. Hallucinated numbers in a board deck are a real harm, and a geometry checker cannot catch them. The team names fact tracing as a next signal, which is where we'd want the effort to go.

Two caveats. The training data is synthetic and generated by a large open teacher, so the model's taste is inherited from that teacher. And PresentBench doesn't measure whether a deck follows a specific design language, which is the headline promise. The authors say so themselves. Until that gap is closed, "your design language" is a claim to test, not a result.

Palette was trained on around 1,000 synthesized decks covering roughly 10,000 slides, each with counterfactual reasoning explaining a rejected alternative. Users can steer output with short bracketed directives in the plan, drawn from 12 categories such as layout, tone, audience and emphasis. For the benchmark runs, a separate frontier model drafted each plan from the source document before Palette built the deck.