AI Weekly Malaysia

IBM Research's Palette fine-tunes a 20B open-weight model to generate slide decks in the IBM Design Language by splitting the job the way humans do: a planner stage decides the story and per-slide content, and a coder stage writes the x/y/width/font-size code for each slide. Training data was synthesized rather than scraped from finished decks, and the model is wrapped in a harness that renders the output and checks it. On PresentBench it ranks first on two of the three domains the team measured; the post does not state that weights, code, or the benchmark are publicly released.

Why: The concrete takeaway is the two-stage planner-plus-coder split with a render-and-check harness, aimed at the failure mode where a small self-hosted model emits overlapping shapes and text that runs off the slide because it does layout arithmetic blind. If confidentiality rules stop you sending deck content to a frontier API, this is a template for what a 20B self-hosted model needs around it — not just fine-tuning, but a verification loop. Note the benchmark is the authors' own and two-of-three domains, so treat it as a design pattern, not a drop-in replacement.