· Nov 3, 2026
Palette: teaching a small model to build decks in your design language
Markdown · JSON · Original source
Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo of IBM Research describe Palette, a 20B for generating slide decks in IBM Design Language. Their approach separates planning from drawing: a designer role makes a structured brief, then a coder role writes code for each slide. The authors say this lets the slide-coding work run in parallel and makes the brief editable. They trained the model on roughly 1,000 synthetic decks, with and examples repaired by a checker. A harness renders slides, checks for overlap and overflow, and keeps a repair only when the checker confirms improvement. In the authors’ PresentBench evaluation, Palette ranked first in Economics and Education, third in Talk, and second across the three domains. The authors note that PresentBench does not measure adherence to a specific design language, and that they know of no that does. The authors present Palette as a way to retain control over design and keep confidential content within an organization’s infrastructure. Their results address deck quality on the reported benchmark, not whether a model reliably follows a company’s design language. How well Palette reproduces other organizations’ templates remains unmeasured; the authors say they are exploring re-skinning from existing .pptx files without retraining.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Nov 3, 2026 (2026-11-03T16:35:36.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-10T19:01:53.712Z. First observed: 2026-11-03T16:36:22.857Z. Retrieved: 2026-11-03T16:46:00.087Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evaluation covers three PresentBench domains and uses plans drafted by a separate from source documents; it does not measure adherence to a specific design language. The source reports that re-skinning output from existing .pptx files without retraining is being explored, not demonstrated.
Open question
How well does Palette reproduce an organization’s design language, given that PresentBench does not measure this and the authors report no that does?
Source-stated
The authors describe Palette as a 20B open-weight model trained to build decks following IBM Design Language.
We fine-tuned a 20B model to build decks the way people do: a designer plans the deck and a coder draws each slide. Its training data follows the IBM Design Language
Source-stated
Palette divides deck creation between a designer role that creates a structured brief and a coder role that generates slide code; slide coding can run in parallel.
The designer reads the plan and writes a structured JSON brief: the palette, the type system, and for each slide its layout, its exact content and a short design rationale.
Source-stated
The authors generated roughly 1,000 decks, with reasoning traces, then rendered and checked examples and repaired broken slides.
We generated around 1000 decks covering roughly 10,000 slides. Every example also carries the reasoning behind it, written as a counterfactual
Source-stated
In the authors’ PresentBench evaluation, Palette ranked first on Economics and Education, third on Talk, and second across the three domains.
Across the three domains it places second, 1.1 points behind NotebookLM (59.7 vs 60.8) and ahead of the rest of the field, as a single 20B open-weight model.
Source-stated
The authors say PresentBench does not measure adherence to a specific design language and that they know of no benchmark that does.
PresentBench measures whether a deck is good, but not whether it follows a specific design language. We're not aware of a benchmark that does
Escape Meridian interpretation
The reported evaluation does not establish how well Palette follows design languages beyond the one used in its training data.
PresentBench measures whether a deck is good, but not whether it follows a specific design language.
Source and contributor context
Hugging Face · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Chain of thought
The step-by-step working a model writes out before giving its final answer.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Editorial provenance
Version 1. Publication ID: 8c5da76e-881d-4fab-8514-80844c95ca77.
Prepared by server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna); grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T19:01:53.712Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b642a1c7068b0aca1f2e14ff2b1b4ce62d0ed606ec486208756218d81bd1db97. Source text hash: 3df5015e8c967830c0d784ccddb2e68799454375c945f40724711fa0c8657305.
Revision and source-check history
Version 1 · 2026-11-10T19:01:53.712Z
Palette: teaching a small model to build decks in your design language
Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo of IBM Research describe Palette, a 20B for generating slide decks in IBM Design Language. Their approach separates planning from drawing: a designer role makes a structured brief, then a coder role writes code for each slide. The authors say this lets the slide-coding work run in parallel and makes the brief editable. They trained the model on roughly 1,000 synthetic decks, with and examples repaired by a checker. A harness renders slides, checks for overlap and overflow, and keeps a repair only when the checker confirms improvement. In the authors’ PresentBench evaluation, Palette ranked first in Economics and Education, third in Talk, and second across the three domains. The authors note that PresentBench does not measure adherence to a specific design language, and that they know of no that does. The authors present Palette as a way to retain control over design and keep confidential content within an organization’s infrastructure. Their results address deck quality on the reported benchmark, not whether a model reliably follows a company’s design language. How well Palette reproduces other organizations’ templates remains unmeasured; the authors say they are exploring re-skinning from existing .pptx files without retraining.
Original source
· Article
Source details
Original source date: Nov 3, 2026 (2026-11-03T16:35:36.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-10T19:01:53.712Z. First observed: 2026-11-03T16:36:22.857Z. Retrieved: 2026-11-03T16:46:00.087Z.
Limits
The evaluation covers three PresentBench domains and uses plans drafted by a separate from source documents; it does not measure adherence to a specific design language. The source reports that re-skinning output from existing .pptx files without retraining is being explored, not demonstrated.
Open question
How well does Palette reproduce an organization’s design language, given that PresentBench does not measure this and the authors report no that does?
Source-stated
The authors describe Palette as a 20B open-weight model trained to build decks following IBM Design Language.
We fine-tuned a 20B model to build decks the way people do: a designer plans the deck and a coder draws each slide. Its training data follows the IBM Design Language
Source-stated
Palette divides deck creation between a designer role that creates a structured brief and a coder role that generates slide code; slide coding can run in parallel.
The designer reads the plan and writes a structured JSON brief: the palette, the type system, and for each slide its layout, its exact content and a short design rationale.
Source-stated
The authors generated roughly 1,000 decks, with reasoning traces, then rendered and checked examples and repaired broken slides.
We generated around 1000 decks covering roughly 10,000 slides. Every example also carries the reasoning behind it, written as a counterfactual
Source-stated
In the authors’ PresentBench evaluation, Palette ranked first on Economics and Education, third on Talk, and second across the three domains.
Across the three domains it places second, 1.1 points behind NotebookLM (59.7 vs 60.8) and ahead of the rest of the field, as a single 20B open-weight model.
Source-stated
The authors say PresentBench does not measure adherence to a specific design language and that they know of no benchmark that does.
PresentBench measures whether a deck is good, but not whether it follows a specific design language. We're not aware of a benchmark that does
Escape Meridian interpretation
The reported evaluation does not establish how well Palette follows design languages beyond the one used in its training data.
PresentBench measures whether a deck is good, but not whether it follows a specific design language.
Source and contributor context
Hugging Face · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Chain of thought
The step-by-step working a model writes out before giving its final answer.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Editorial provenance
Version 1. Publication ID: 8c5da76e-881d-4fab-8514-80844c95ca77.
Prepared by server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna); grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T19:01:53.712Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b642a1c7068b0aca1f2e14ff2b1b4ce62d0ed606ec486208756218d81bd1db97. Source text hash: 3df5015e8c967830c0d784ccddb2e68799454375c945f40724711fa0c8657305.
passed · 2026-11-10T19:01:53.712Z · anthropic/claude-opus-5.5
Publication: 8c5da76e-881d-4fab-8514-80844c95ca77
attribution — pass: The reading correctly attributes Palette and its claims to Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo of IBM Research, who are listed as the authors of this Hugging Face post. Performance claims are kept as the authors' own reported results.
Praveen Venkateswaran praveenv ibm-research Anupama Murthi anupamamurthi ibm-research Hamid Adebayo
summary — pass: The summary faithfully conveys the authors' bet: a small model that an organization runs itself, built around a designer/coder split, synthetic training data, and a check-and-repair harness. It preserves the mixed benchmark ranking and the authors' admission that design-language adherence is unmeasured. The central ideas are readable without specialist background. Terms such as open-weight and 20B are not explained, but the surrounding context about owning and self-hosting the model keeps the meaning clear. The question follows from a gap the source itself states and is not a leading claim.
Palette tests one bet: that a small open-weight model, trained on synthetic data in your design language and checked by a harness, can compete with frontier systems while you keep ownership of the result.
date — pass: The originalPublished value of 2026-11-03 matches the publication date shown on the post. There is no indication of an earlier submission or first publication.
Published November 3, 2026
support — pass: Every claim is grounded in the source text. This includes the 20B model trained on data following the IBM Design Language, the designer/coder split with parallel slide coding, the roughly 1,000 synthetic decks, and the rule that a repair is kept only when the checker confirms improvement. It also covers the PresentBench rankings and the stated benchmark gap. The interpretation claim about untested generalization is labeled and conditional. The limitations accurately note that plans were drafted by a separate frontier model and that re-skinning is still being explored.
The key rule is that a rewrite is kept only if the checker confirms the slide improved; otherwise the original stays.
media — pass: The selected media is the verified direct link to the original Hugging Face article. It is the most direct way for a reader to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceThe step-by-step working a model writes out before giving its final answer.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceThe step-by-step working a model writes out before giving its final answer.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference