RULER matches DeepSeek-V3 on SVG using text-only rubrics

RULER, a training method for text-to-SVG models, matches the substantially larger DeepSeek-V3 on two benchmarks without any paired ground-truth images or human preference labels. The arxiv preprint from Hangyu Ran and colleagues starts from a specific bind: SVG generation from natural language "is an open-ended task without absolute visual ground truth," so scalar metrics trained on photographs give no faithful reward signal.

The workaround is to derive a per-instance rubric from the prompt itself. Each instruction is converted into six rubric items spanning semantic, visual, and stylistic dimensions. A vision-language model scores rendered outputs item-by-item, and the weighted satisfactions feed Group Relative Policy Optimization. Rubric scores climbed from 0.432 and 0.395 to 0.693 and 0.683 on MMSVG-Illustration and MMSVG-Icon respectively.

The authors report the trained model's outputs "surpass dedicated SVG specialists and match the substantially larger DeepSeek-V3," and that rubric prompting tracks human judgment more tightly than scalar metrics "across samples and within instructions." The abstract does not name the VLM used as scorer, and it publishes no head-to-head win rate against DeepSeek-V3 beyond the aggregate rubric numbers.