Palette is a presentation generation model derived from a 20B open-weight model, meaning its weights are accessible and it can run on your own infrastructure. Its authors trained it to produce slides in IBM‘s design language. On the PresentBench benchmark, it ranks first in two of the three measured domains and second on the average across the three domains, behind NotebookLM.
Asking a large language model for a presentation usually yields a correct result, but the outputs end up looking similar from one query to the next, with a recognizable house style. For a company that wants to adhere to its own presentation template, or that must keep confidential content on its own servers, this approach quickly hits its limits. Small open-weight models, on the other hand, often produce defective slides: titles lost in the background, overflowing text, overlapping shapes.
A designer and a coder
The problem stems from three difficulties. The model cannot see the slide it is building: it writes code that places each element at a given position, width, and font size, and only discovers its errors after rendering. It must also make two types of decisions simultaneously, editorial (what to say, which chart to choose) and geometric (where to place elements). Finally, training data was lacking: finished presentations exist, but not the reasoning behind the design choices or the code that produced them.
Palette adopts a human-like method. A designer outlines the plan and writes a structured brief in JSON format, specifying the palette, typographic system, and, for each slide, the layout, content, and a brief justification. A coder then writes, for each slide, the pptxgenjs code that draws it, preceded by their reasoning. Both roles are handled by a single LoRA adapter, a small training layer added to the base model without modifying its weights. Since the coder’s calls are launched in parallel, one presentation per slide, the coding step would, according to the authors, be approximately 6 to 7 times faster.
Synthetic data and a controller
Lacking a public corpus, the team generated around 1,000 presentations, or nearly 10,000 slides, using an open-source large “teacher” model. Each example includes a counterfactual reasoning, mentioning an alternative and explaining why it was rejected. When a generated slide has a defect, it is corrected, and its reasoning is rewritten accordingly.
To ensure fidelity to the source content, the authors found that a rule in the prompt was insufficient. What worked was correcting the reference presentations provided to the teacher, which copies them much more than it follows instructions. Users can also steer the result with directives written in double brackets, such as “[]”. These directives, divided into twelve categories, are phrased in multiple ways in the training data.
Before returning a presentation, Palette goes through a verification harness. Simple rules correct known errors, then each slide is rendered and measured by a geometric controller that flags overlapping, truncated, or out-of-bounds text. Flagged slides are rewritten by a separate editor model, but the rewrite is only kept if the controller confirms the improvement. On 131 presentations from PresentBench, the harness adds an average of 3.5 points.
Mixed results
PresentBench evaluates five criteria: fundamentals, visual design, completeness, accuracy, and fidelity. Palette achieves its highest ranking in Economics and Education. Conversely, in the Talk domain, where presentations are narrative, two closed systems outperform it and it ranks third. The average across the three domains places it at 59.7, compared to 60.8 for NotebookLM.
The authors indicate that next steps include using the harness controls as reward signals for GRPO-style reinforcement learning, as well as extracting figures from source documents with Docling.
