RULER.

RULER: Instance-aware Rubric Rewards for SVG Generation

Hangyu Ran1,2,*Yuhao Zheng3,*Yingying Zhang1Kevin Qinghong Lin4,†Han Peng1,†
1 Ant Group2 The Hong Kong University of Science and Technology (Guangzhou)3 Independent Researcher4 University of Oxford

* Equal contribution † Corresponding authors

arχivPaper Daily Paper Code Coming soon
GENERATED WITH RULERSelected outputs from Figure 5
01 /RULER-generated eye icon with eyelashes and a blue iris
A stylized eye icon
with eyelashes
02 /RULER-generated golden hexagonal beehive pattern
A hexagonal
beehive pattern
03 /RULER-generated compass needle pointing northeast
A compass needle
pointing northeast
04 /RULER-generated orange and yellow starburst explosion
A starburst
explosion icon
05 /RULER-generated target with concentric rings
A target with
concentric rings
Explore the qualitative comparison

01 / OVERVIEW

Abstract.

Instance-aware.
Multi-axis.
Ground-truth-free.

Generating Scalable Vector Graphics (SVG) code from natural-language instructions is an open-ended task without absolute visual ground truth, leaving both evaluation and policy optimization without a faithful signal. Scalar metrics (CLIP, Aesthetic) calibrated on natural images transfer poorly to stylized vector content, and reusing them as RL rewards triggers reward hacking. We address both limitations with rubric-based scoring.

We first establish empirically that prompting a vision–language judge with a multi-axis rubric correlates with human judgments far better than scalar metrics, both across samples and within instructions. Building on this finding, we introduce RULER (Instance-aware Rubric Rewards for Reinforcement LEaRning), which converts each instruction into an instance-aware rubric of six items spanning semantic, visual, and stylistic axes; a judge VLM scores rendered rollouts item-by-item, and the weighted satisfactions form a fine-grained reward optimized via Group Relative Policy Optimization.

Because the rubric is derived from text alone, RULER requires neither paired SVG ground truth nor human preference labels. On MMSVG-Illustration and MMSVG-Icon, RULER lifts the rubric score from 0.432/0.395 to 0.693/0.683, surpassing dedicated SVG specialists and matching the substantially larger DeepSeek-V3, with ablations identifying rubric design as the active lever for RL on open-ended SVG generation.

02 / MOTIVATION

Rubric-based evaluation
for SVG quality.

Standard scalar metrics (CLIP and Aesthetic) mis-rank or fail to penalize a broken SVG, whereas our Rubric score aligns with human preference.

FIGURE 01Evaluation metrics & instance-aware rubric
Figure 1. Comparison of evaluation metrics and the RULER rubric.
TABLE 01Comparison of reward paradigms
Table 1. Comparison of reward paradigms for SVG generation. We evaluate existing approaches across five critical desiderata. GT-free denotes that ground-truth visual references are not required.

03 / THE METHOD

Instance-aware rubrics.
Fine-grained rewards.

Because the rubric is derived from text alone, RULER requires neither paired SVG ground truth nor human preference labels.

01—02

Semantic Fidelity

High-level concept readability and the visual presence of major components, distinctive cues, and prompt-specific relations.

03—04

Visual Quality

Silhouette and form refinement together with composition and canvas design.

05—06

Rendering Style

Rendering finish and execution cleanliness coupled with style cohesion and designed visual interest.

FIGURE 02The RULER framework
Figure 2. Overview of RULER. A frontier model derives an instance-aware rubric—six items across semantic, visual, and stylistic axes—from each text instruction (left). The policy samples SVG rollouts that are rendered and scored by a judge VLM against the rubric; the weighted per-item satisfactions form the GRPO reward signal that drives policy updates (right).

04 / EXPERIMENTS

Results, at a glance.

RULER outperforms every baseline family on the primary Rubric metric across both benchmarks.

MMSVG-Illustration
0.693↗
Rubric score from 0.432Table 2 · Qwen3-8B → RULER
MMSVG-Icon
0.683↗
Rubric score from 0.395Table 2 · Qwen3-8B → RULER
Human score correlation
0.7929
Spearman’s ρFigure 3 · 900 SVG samples
Pairwise ranking agreement
0.7574
Goodman–Kruskal γFigure 3 · 300 evaluation triples
TABLE 02MMSVG-Illustration & MMSVG-Icon
Table 2. Main results on MMSVG benchmarks. Our method achieves the best universal rubric scores on both benchmarks while maintaining competitive CLIP and HPS performance with significantly fewer tokens than optimization-based methods.

To capture holistic visual quality, we additionally introduce a Rubric score that prompts GPT-5-mini (OpenAI, 2025) as an independent VLM-as-Judge to rate each rendered SVG against a shared universal rubric; we treat this Rubric score as the primary indicator of overall quality.

Blinded human preference

As shown in Table 3, RULER achieves a non-tie win rate above 50% against every evaluated baseline, ranging from 53.3% against VectorFusion to 96.5% against JanusCoder.

TABLE 03150 prompts · Win rates exclude ties
Table 3. Blinded human preference results. Win rates compare RULER against each baseline on 150 MMSVG-Bench prompts and exclude ties.

Reward design analysis

To isolate the contribution of our reward design, we fix the base model (Qwen3-8B) and the GRPO optimizer, and vary only the reward signal across four configurations.

TABLE 04Comparison of RL reward designs
Table 4. Comparison of different RL reward designs on MMSVG benchmarks. C, A and H denote CLIP, Aesthetic and HPS, respectively.

Across models and rubric generators

To examine whether the effectiveness of RULER depends on a particular base model or rubric generator, we evaluate the framework across different base-model scales and rubric sources while keeping the remaining training setup unchanged.

TABLE 05Base-model scales · Mean ± standard deviation over five runs
Table 5. Robustness across base-model scales on MMSVG benchmarks. We compare RULER-4B and RULER-8B with their corresponding base models and representative baselines. Results are reported as mean ± standard deviation over five runs.
TABLE 06Rubric generators · MMSVG-Icon
Table 6. Robustness across rubric generators on MMSVG-Icon. RULER uses the same Qwen3-8B policy with different rubric generators. Results are reported as mean ± standard deviation over five runs.

EVALUATOR ALIGNMENT

Closer to human judgment.

Figure 3. Rubric score shows superior human alignment compared with Aesthetic and CLIP.

RUBRIC ABLATION

Every axis contributes.

Figure 4. Ablation of rubric design on MMSVG-Illustration and MMSVG-Icon.

05 / QUALITATIVE COMPARISON

From instructions
to vector graphics.

Figure 5 presents qualitative comparisons on representative icon-style prompts. Compared with the baselines, RULER generates SVGs that are more prompt-faithful, visually expressive, and stylistically polished.

FIGURE 05RULER & four baselines
Figure 5. Qualitative comparison of scalable vector graphics generation between RULER and four baselines.

06 / REFERENCE

Citation.

If you find this work useful,
please cite our paper.

citation.bib
@misc{ran2026rulerinstanceawarerubricrewards,
  title={RULER: Instance-aware Rubric Rewards for SVG Generation},
  author={Hangyu Ran and Yuhao Zheng and Yingying Zhang and Kevin Qinghong Lin and Han Peng},
  year={2026},
  eprint={2609.25270},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.25270},
}
arXiv:2609.25270

Paper figure

Open original ↗