What happened
In a post published November 3, 2026, IBM Research presents Palette as an experiment in building presentation decks with a model an organization can own and run on its own infrastructure. The team’s stated problem is that frontier systems may converge on their own visual defaults, while smaller self-hostable models often produce broken layouts. Palette fine-tunes a 20B open-weight base with a LoRA adapter, rather than conducting full fine-tuning, to teach the model a deck-building format and skill aligned with the IBM Design Language.
The workflow separates two kinds of work that are usually entangled in a single generation pass. A “designer” reads a plan and produces a structured JSON brief containing the palette, type system, slide layouts, exact content, and short design rationales. A “coder” receives one slide specification and writes the pptxgenjs JavaScript that draws it, preceded by reasoning about the placement. The brief is meant to be readable and editable: a person or another agent can change one slide’s layout without regenerating the entire deck. Since the specifications are fixed before drawing begins, the coding calls can run in parallel, one per slide. IBM reports this makes that stage roughly six to seven times faster.
IBM says no public corpus contained both deck plans and the design decisions and code behind them, so it synthesized its own training data with a large open teacher model. Recipes specified genre, length, layouts, and user directives; plans ranged from structured outlines to raw content dumps. The resulting corpus contained about 1,000 decks and roughly 10,000 slides. Examples included counterfactual reasoning: not just a selected chart or layout, but an alternative that was considered and why it was rejected. The team found coder examples were around five times longer than designer examples, then oversampled designer examples fivefold to avoid undertraining that role.
The generation path includes a check-and-repair harness. It first corrects known simple errors, then renders each slide and uses geometry checks to flag overlaps, clipping, and content running outside the slide. A separate editor model rewrites code for flagged slides. Crucially, a rewrite is retained only if the checker verifies that it improved the slide; otherwise Palette keeps the original. IBM used the same kind of checking while producing synthetic examples, repairing broken slides and rewriting their reasoning. Across 131 PresentBench decks, IBM reports that adding the harness increased scores by 3.5 points on average.
For the reported evaluation, a separate frontier model first converted each PresentBench source document into a plan, and Palette then built the deck from that plan. PresentBench scores fundamentals, visual design, completeness, correctness, and fidelity. Palette ranked first in Economics and Education, third in Talk, and second across the three-domain mean, with 59.7 against NotebookLM’s 60.8. IBM says Palette performs best where structured content and numeric accuracy matter; two closed systems did better in the narrative, speech-oriented Talk domain. The post does not announce public access, pricing, a model release, or a license.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
Palette supports design-systems-become-agent-context. Its workflow turns palette, typography, layouts, content constraints, and directives into an explicit structured brief that a model can consume and a person can edit. The evidence remains narrow: IBM has shown one research workflow and has not measured whether it transfers faithfully across organizations’ design systems.
Connections to other events
The archive includes earlier Telegram coverage of Paper2Slides, a pipeline for turning papers and documents into slides, and open-slide, a React-based runtime for agent-generated presentations. Palette addresses the same broad output category but shifts attention to a trainable planning-and-code workflow, a specific design language, and a conservative render-repair gate.
Why it matters
The central practical contribution is not that a model can make slides. Palette makes the intermediate representation inspectable, then puts an acceptance gate between generation and delivery. Teams can intervene at several points: revise the plan, alter a slide brief, inspect generated drawing code, or decline a repair that fails to improve measurable output.
That is a reusable pattern for structured AI output. If a team can define cheap checks for visible or geometric failures, it can allow automated repair without blindly trusting a second model pass. Parallelizing independently specified slides also offers a concrete way to reduce latency. The work is relevant to charts, diagrams, and UI layouts, where an artifact can be rendered and examined rather than judged only by prose quality.
Limits and uncertainty
The source is IBM Research’s account of its own system and results. It gives no public repository, deployment instructions, access terms, pricing, or license for Palette. PresentBench did not measure the claim most relevant to enterprises with strict brand requirements: whether a generated deck faithfully follows a specific design language. The pipeline also used a separate frontier model to create evaluation plans, so the result is not a document-to-deck measurement of Palette alone. Results from 131 decks across three domains do not establish reliability for other templates, document types, or confidential enterprise material. Reinforcement learning from checker signals, figure extraction, and template re-skinning are described as future work.
Skill to build now
Pick one structured artifact your team generates repeatedly, such as a chart, dashboard panel, or slide. Define three machine-checkable failure conditions this week: for example, text clipping, overlapping boxes, and a missing required label. Render outputs automatically, record those checks beside the generation, and only accept a repair when it improves the measured result. This creates a verification boundary before investing in fine-tuning.
Reader takeaway
Palette is evidence for a workflow, not yet a drop-in presentation product: use an editable plan, generate bounded artifacts independently, and verify visible output before delivery. Its scores are promising, but IBM has not shown that the system reliably reproduces a customer’s own design language—the capability the project is ultimately meant to enable.