AI BriefingIBM Research Palette: 20B open-weight model ranks second on PresentBench, first in Economics and Education
·2026.11.04 01:35
Key point
A fine-tuned 20B model using a designer-coder split and check-repair harness achieves a 59.7 mean score on PresentBench, ranking second overall and first in Economics and Education.
Details
- IBM Research released Palette, a 20B open-weight model fine-tuned with LoRA to generate slide decks in specific design languages, such as the IBM Design Language.
- On the PresentBench benchmark, Palette ranks first in the Economics and Education domains and third in the Talk domain, achieving a three-domain mean score of 59.7, which is 1.1 points behind the leading system, NotebookLM (60.8).
- The system splits generation into a designer role that creates a structured JSON brief and a coder role that writes pptxgenjs JavaScript, allowing parallel slide rendering that is 6–7 times faster than sequential generation.
- A check-and-repair harness validates every slide for overlaps and overflow, adding an average of 3.5 points to the score across 131 test decks by keeping only repairs that measurably improve the output.
- Training data was synthesized using a large open teacher model to generate approximately 1,000 decks (10,000 slides) with counterfactual reasoning traces, addressing the lack of public datasets for deck generation rationale.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
A $500 RL fine-tune of a 9B open model beat frontier models on catalog review
Find Hidden Misalignment in Training Data
Toss Reduces LLM Costs with Selective Neural Approximation (SNAP)