IBM Research describes Palette, a fine-tuned 20-billion-parameter model designed to create slide decks in IBM’s design language. Its reported benchmark results put it first in two domains and second overall, with a practical twist: a checking system can catch and repair some layout mistakes before the deck reaches you.
Watch Desk analysis
What happened
Palette splits the job between two roles. A “designer” turns a plan into a structured brief covering content and layout; a “coder” writes the code for each slide. The slides can then be generated in parallel, which IBM says makes that stage six to seven times faster.
IBM says it trained the model using roughly 1,000 synthetic decks, covering about 10,000 slides, and checked those examples for layout faults. In tests on 131 PresentBench decks, its harness improved scores by an average of 3.5 points. Across three benchmark domains, Palette ranked first in Economics and Education, third in Talk and second overall, 1.1 points behind NotebookLM. These are IBM’s reported results; the article notes that the benchmark does not test how well a deck follows a particular design language.
Why it matters
Slide generation often fails in mundane, visible ways: text overflows, shapes collide, and a model cheerfully invents missing details. Palette’s approach tackles some of that with a separate checker that keeps a repair only when the slide measurably improves. IBM says the model also uses reference decks to improve faithfulness, rather than relying on prompt rules alone.
The result points to a useful path for AI tools that make structured work: divide the task, check the output, and reject fixes that make things worse. That could matter beyond presentations, though the reported benchmark results do not establish performance on other tasks.
Our read
The most interesting part is not that a model can write slide code. It is the modest, sensible machinery around it: separate editorial choices from geometry, then make repairs earn their place. That is a better pitch than asking users to trust a polished-looking deck on sight.
IBM presents Palette as a research project, not a finished product for general use. And while it performs well on the reported benchmark, it still trails the leader overall and places third on narrative-driven Talk decks. Promising work, with a clear next test: whether it can follow a company’s actual templates without retraining.
What to watch
- Whether IBM releases the model, training details or tools for others to reproduce the results.
- How Palette performs on decks built from real company templates and confidential material.
- Whether the checking-and-repair approach transfers to other structured outputs, such as charts or interface layouts.
Discussion spark: For AI-generated work documents, would you trust a smaller model with a reliable checker over a stronger model that produces better-looking first drafts?
Sources and evidence
- Palette: teaching a small model to build decks in your design language (3 November 2026, 16:35 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.