RULER: Instance-aware Rubric Rewards for SVG Generation
RULER turns each SVG prompt into a six-part grading rubric, then uses that rubric as the reward signal for reinforcement learning.
The paper argues that CLIP and aesthetic scores do not transfer well to stylized SVGs and can be gamed when used as rewards. Its method has a vision-language judge score rendered SVG outputs item by item across semantic, visual, and stylistic criteria. Because the rubric is generated from the text prompt, the approach does not need paired SVG ground truth or human preference labels. Reported scores rise from 0.432/0.395 to 0.693/0.683 on MMSVG-Illustration and MMSVG-Icon. HF Daily Papers' note
The paper argues that CLIP and aesthetic scores do not transfer well to stylized SVGs and can be gamed when used as rewards. Its method has a vision-language judge score rendered SVG outputs item by item across semantic, visual, and stylistic criteria. Because the rubric is generated from the text prompt, the approach does not need paired SVG ground truth or human preference labels. Reported scores rise from 0.432/0.395 to 0.693/0.683 on MMSVG-Illustration and MMSVG-Icon. HF Daily Papers' note
score 5