What's this about?
RULER derives six criteria from each request, renders candidate SVGs and asks a vision-language judge to score each criterion for RL.
Why not use one image score?
A scalar can reward superficial resemblance and invite reward hacking. A rubric separates semantic correctness, composition and style, though the judge can still carry biases.
Why does this matter?
Author-reported preprint; 13 HF upvotes.
- It needs neither paired SVGs nor human preference labels.
- Reported rubric scores rise from 0.432/0.395 to 0.693/0.683.
- Visual quality still needs evaluation beyond the training judge.