IBM Research's Palette: a 20B open model builds on-brand slide decks

IBM Research says a 20-billion-parameter open-weight model, fine-tuned to build slide decks in IBM's design language, ranks second across three domains of the PresentBench benchmark. The result is in a post published on 3 November 2026 on the Hugging Face blog by Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo. The system is called Palette.

The problem it targets

The authors say frontier models write good decks but drift back to their own house style however you prompt them, and some content cannot leave your systems at all. Small self-hostable models, they report, produce broken slides: titles that vanish into the background, text running off the edge, shapes stacked on each other. Palette tests whether training on your own design language plus a checking harness closes that gap.

How it is built

One LoRA adapter serves two roles, according to the post. A designer reads the plan and writes a structured JSON brief with the palette, type system and, for each slide, its layout, exact content and a short rationale. A coder takes one slide's spec and writes pptxgenjs JavaScript that draws it. Because the brief fixes each slide's context first, the coder calls run in parallel, which the authors say makes that stage about 6-7 times faster.

The training data was synthetic: around 1,000 decks covering roughly 10,000 slides, generated by a large open teacher model. Each example carries reasoning written as a counterfactual, such as why a table was rejected in favour of a bar chart. Every deck was rendered and run through the same overlap and overflow checker used at inference; broken slides were repaired and their reasoning rewritten.

The harness

Rules first fix known mistakes, such as a malformed colour code that turns every slide black. Each slide is then rendered and measured, and flagged slides go to a separate editor model. The authors' key rule is that a rewrite is kept only if the checker confirms the slide improved, so the harness can fix a slide but cannot make one worse. Across 131 PresentBench decks they report it adds 3.5 points on average.

What the numbers say

On PresentBench, which scores decks on fundamentals, visual design, completeness, correctness and fidelity, the post reports Palette first on Economics and Education, third on Talk, and second on the three-domain mean: 59.7 against NotebookLM's 60.8. Baselines come from the PresentBench paper. For these runs a separate frontier model drafted each plan from the source document, so the result is not a single-model pipeline.

What it leaves out

• No cost or latency figures per deck, and no hardware for the 20B model.
• No test of the stated goal: the authors say they know of no benchmark that measures whether a deck follows a specific design language, so on-brand fidelity is not scored.
• Figure-heavy decks are a stated weakness; the authors list figure extraction as future work.
• The ablation reports a mean gain only; the spread across the 131 decks is not given.

What a team can take from it

The transferable part is process, not the model. The authors write that fixing a faulty reference deck changed the teacher's output far more reliably than adding a prompt rule, and that building the checker first made both the training data and the outputs trustworthy. Any task whose output has structure you can verify in code, such as charts or UI layouts, fits the same loop.