Checklist rewards beat similarity scores for vector graphics, paper reports
single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC
What this means for you
Nothing to act on yet: this is a method paper, with no released code, weights or API, and no evaluation date attached to the numbers. If you build generative design tooling, the transferable idea is using a rubric as the training reward — but you would have to reimplement it and check the judge against your own reviewers first.
An arXiv paper describes RULER, a method that scores generated SVG code — images stored as code rather than pixels — against a six-item checklist derived from the text prompt alone. A vision-language model judges each rendered candidate item by item, and those scores become the reward for reinforcement learning. On two MMSVG benchmarks the authors report rubric scores rising from 0.432 and 0.395 to 0.693 and 0.683, beating dedicated SVG systems and matching DeepSeek-V3. Scalar metrics such as CLIP and aesthetic scores, they say, transfer poorly to stylised vector art. No evaluation date, code release or human baseline is given.
Models in this story
Key facts
- ·RULER turns each text instruction into a six-item rubric spanning semantic, visual and stylistic axes, derived from text alone. source
- ·A judge vision-language model scores rendered rollouts item by item, and the weighted satisfactions are optimised with Group Relative Policy Optimization. source
- ·The method requires neither paired SVG ground truth nor human preference labels. source
- ·Reported rubric scores rise from 0.432 to 0.693 on MMSVG-Illustration and from 0.395 to 0.683 on MMSVG-Icon. source
- ·The authors report surpassing dedicated SVG specialists and matching DeepSeek-V3, which they describe as substantially larger. source
- ·The paper is arXiv 2609.25270, posted 20 September 2026, with no evaluation date given for the scores. source
What the sources say
- Hugging Face Daily Papers (research) — Preprint using a judge model's per-instruction rubric scores as the reinforcement-learning reward for vector output
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersRULER: Instance-aware Rubric Rewards for SVG Generation2026-09-20