#

Key takeaways

• Palette is a deck-building system described by IBM Research on November 3, 2026.

• It fine-tunes a twenty-billion-parameter model for presentation decks.

• A single LoRA adapter supports both brief writing and slide code generation.

• The system checks slides for formatting, overlap, clipping, and overflow before returning a deck.

• The authors report gains from the repair harness and note future directions.

IBM Palette and its deck-building workflow

IBM Research described Palette in an article on Hugging Face. The project focuses on creating presentation decks that follow IBM Design Language. It fine-tunes a twenty-billion-parameter model for this task. The system uses one LoRA adapter for two roles. A designer writes a structured JSON brief. A coder then generates pptxgenjs JavaScript for individual slides.

The brief sets slide context first. That design choice allows coding calls to run in parallel. The source presents this as part of the workflow, not as a separate product claim. The emphasis is on structured input and slide-level generation. The result is a pipeline that links planning and code creation.

Training data and method

The researchers report generating around one thousand training decks with roughly ten thousand slides. The source is explicit that this describes the training method. It is not evidence that those slides represent real business activity. The report also says the team values factual faithfulness to the supplied brief. The authors report that reference examples helped improve faithfulness to the supplied input.

The researchers also say that improving reference examples helped more than adding prompt rules. This is an important detail because it points to the training setup, not just the prompt format. The source does not present this as a universal rule. It is a reported result from the Palette work. The main takeaway is that example quality mattered more than extra instructions in their process.

Checking, repair, and slide validation

Palette uses a checking and repair harness before returning a deck. The harness fixes known formatting errors. It also renders slides and checks for overlap, clipping, and overflow. These checks focus on presentation quality and layout integrity. They are part of the system's operational flow.

When the checker flags slide code, an editor rewrites it. The rewrite is kept only if the checker measures improvement. That creates a feedback loop between generation and validation. The source frames this as a controlled repair step. It does not describe the harness as fully autonomous. It describes a system that evaluates and accepts changes only when they improve the measured result.

Across 131 PresentBench decks, the authors report an average gain of 3.5 points from the harness. This is a reported evaluation result from the source material. It supports the claim that the repair step improved outcomes in that benchmark setting. The source does not extend that result beyond the reported evaluation.

Evaluation results and comparison

In three evaluated domains, Palette ranks first in Economics and Education and third in Talk. Its three-domain average is 59.7. NotebookLM's score is 60.8. The source presents these numbers as part of the evaluation summary. It also notes that a separate frontier model drafted input plans for these evaluations, and that baseline scores came from the PresentBench paper.

The source adds an important caution about the benchmark itself. PresentBench does not measure adherence to a particular design language. That means the benchmark results do not directly test whether a deck matches IBM Design Language. The evaluation therefore covers performance on the benchmark tasks, while the design-language goal remains a separate system objective.

Future directions and operational considerations

The researchers describe reinforcement learning from checks, figure extraction, and adapting existing templates as future directions. The source clearly labels these as future work, not completed features. That distinction matters because it separates current system behavior from planned extensions. The current description centers on fine-tuning, structured briefs, slide code generation, and repair checks.

From an operational perspective, the source suggests a workflow built around validation. The system does not simply generate slides and stop. It checks output, rewrites flagged code, and keeps changes only when they improve the measured result. That approach can help control formatting issues and layout errors. It also adds a review step that shapes the final deck.

Overall, Palette is presented as a structured deck-generation system with a strong emphasis on faithfulness, layout checking, and measurable repair. The source focuses on how the pipeline works and how it performed in the reported evaluation. It does not claim that the future directions are already implemented. It also keeps the benchmark results separate from the design-language objective.