IBM Research described Palette on November 3, an approach that teaches a small open-weight model to build slide decks in a specific company’s design language. According to the write-up on Hugging Face, Palette uses a single LoRA adapter on a 20-billion-parameter open-weight model and scores 59.7 across three domains of PresentBench, 1.1 points behind Google’s NotebookLM at 60.8. The interesting part is less the score than the recipe: a designer-then-coder split, a checking harness and parallel slide generation, which show how AI presentation tools can be made brand-aware without a giant model.
What Palette is
Palette targets a practical problem. Generic AI deck tools produce plausible slides, but companies want output that follows their own templates, fonts and layout rules. IBM’s team trained a model to follow the IBM Design Language. The base model is a 20B open-weight model that the post does not name, and the training is a single LoRA adapter, a small set of added weights, rather than full fine-tuning. The post does not say whether the weights are released or under what license.
One adapter, two jobs
The adapter plays two roles. As a designer, it decides the structure and look of each slide, including reasoning about rejected alternatives. As a coder, it writes pptxgenjs JavaScript, a library that produces PowerPoint files, one slide at a time and in parallel. Because slides are written in parallel, IBM reports the coding stage is about six to seven times faster than writing the deck sequentially.
Training data was synthetic: about 1,000 decks covering roughly 10,000 slides, generated with a large open teacher model. Designer examples were oversampled five times because coder examples are about five times longer. Users can steer output with directives in double brackets, such as [[make this a timeline]]; the model was trained on 12 directive categories.
The harness that checks the slides
A second ingredient is a harness around the model. It applies rule-based fixes, renders each slide, and runs a geometry checker for overlapping or overflowing elements. Flagged slides go to an editor model, and a rewrite is kept only if the checker confirms an improvement. Across 131 decks, IBM says the harness adds 3.5 points on average. This design, in which a program verifies layout rather than trusting the model, echoes the verifier-based approach we explain in our guide to reinforcement learning with verifiable rewards.
How the results compare
On PresentBench, IBM reports Palette ranked first on the Economics and Education domains, third on Talk, and second on the three-domain mean at 59.7. NotebookLM is first at 60.8. These are IBM’s own reported numbers; we have not rerun the benchmark. The PresentBench dataset is public on Hugging Face, so others can check the results.
For readers choosing a tool today, Palette is a research approach, not a product you can sign up for. Commercial options such as Gamma are available now and are built for general use rather than one company’s template.
Limits IBM acknowledges
IBM lists several gaps. Figure-heavy decks remain an open problem. Talk-style narrative decks score lower. Faithfulness is difficult: the model can invent content when the input is thin, so users should check facts on every slide. Reinforcement learning with checker-based rewards is planned but not yet done.
What this means for teams making decks
Three lessons carry over. First, brand consistency can come from a small adapter trained on examples of your own templates, not only from the largest models. Second, generating slides independently in parallel saves time but needs a final pass for deck-wide coherence. Third, automated geometry checks catch layout errors cheaply. Anyone evaluating an AI deck generator should ask how it handles overflow, templates and source fidelity.
What to watch next
- Whether IBM releases Palette weights or the training recipe.
- Results after IBM adds checker-based reinforcement learning.
- Independent PresentBench runs of Palette and rival tools.
- Support for figure-heavy and chart-driven decks.
Key takeaways
- Palette adapts a 20B open-weight model with one LoRA adapter to follow the IBM Design Language.
- It scores 59.7 on PresentBench’s three-domain mean, versus 60.8 for NotebookLM.
- A render-and-check harness adds 3.5 points on average across 131 decks.
- Release status, license and figure-heavy decks are open questions.
Frequently asked questions
What is Palette from IBM Research?
Palette is an approach that trains a small open-weight model to build slide decks in a specific design language, using a LoRA adapter that acts as both designer and coder.
How does Palette score on PresentBench?
IBM reports 59.7 on the three-domain mean, second place and 1.1 points behind NotebookLM’s 60.8.
Can I download Palette?
The post we reviewed does not state whether the weights are released or under what license, so availability is unclear.
