RULER: Instance-aware Rubric Rewards for SVG Generation

RULER. Instance-aware Rubric Rewards for SVG Generation. By Hangyu Ran and four others, from Ant Group, the Hong Kong University of Science and Technology, Guangzhou, and two other institutions. Published on the twenty-first of September. No code repository is linked to it.

Generating Scalable Vector Graphics, or SVG, code from natural-language instructions is an open-ended task with no absolute visual ground truth. That leaves both evaluation and policy optimization without a faithful signal. Scalar metrics such as CLIP and Aesthetic were calibrated on natural images and transfer poorly to stylized vector content. Reusing them as reinforcement learning rewards triggers reward hacking: the policy learns to please the metric rather than the instruction. So the same weakness damages two things at once: the numbers used to compare systems, and the signal used to train them.

The answer to both problems is rubric-based scoring. The authors first establish empirically that prompting a vision-language judge with a multi-axis rubric correlates with human judgments far better than scalar metrics do. That holds both across samples and within a single instruction. Building on that, RULER, for Instance-aware Rubric Rewards for Reinforcement Learning, turns the rubric into a reward.

Each instruction is converted into its own rubric of six items, spanning semantic, visual and stylistic axes. A judge VLM scores the rendered rollouts item by item. The weighted satisfactions form a fine-grained reward, which is optimized with Group Relative Policy Optimization. Because the rubric comes from the text alone, RULER needs neither paired SVG ground truth nor human preference labels. That is what makes it usable on an open-ended task. There is no reference drawing to compare against, but there is always the instruction, and the rubric is built from that.

On MMSVG-Illustration, RULER lifts the rubric score from zero point four three two to zero point six nine three. On MMSVG-Icon, it goes from zero point three nine five to zero point six eight three. That surpasses dedicated SVG specialists and matches the substantially larger DeepSeek-V3. Ablations identify rubric design as the active lever for reinforcement learning on open-ended SVG generation. The gain comes from what the rubric asks, not merely from having a judge in the loop.

The move is to replace one number, borrowed from natural images, with a short checklist written for each instruction, and then to train against that checklist directly. Each item can be judged on its own, and the checklist needs nothing but the instruction to build.