Literature
Display
← All papers

RULER: Instance-aware Rubric Rewards for SVG Generation

Hangyu Ran · Yuhao Zheng · Yingying Zhang · Kevin Qinghong Lin · Han Peng · * Equal contribution Corresponding authors

· arXiv (preprint) ↗

Language models10 min text · CondensedOriginal paper ↗
Contents

Overview

At a glance

RULER trains a language model to produce SVG code using a rubric tailored to each prompt. A frozen visual judge scores six criteria spanning meaning, visual quality, and rendering style; their weighted average rewards reinforcement learning. On 300 illustration and icon prompts, the trained Qwen3-8B improves the separate evaluation judge’s scores from 0.432/0.395 to 0.693/0.683. Human comparisons favor it over several baselines, although its 53.3% win rate against VectorFusion is a narrow observed edge. Tests stay within one benchmark family, and rubric design can still induce reward hacking.

1 Introduction

Many different vector drawings can satisfy the same instruction, so matching a single reference image is an unsuitable training target. Generic image metrics can also reward attractive but irrelevant graphics. RULER turns a prompt into explicit visual criteria, scores rendered candidates against them, and uses those scores to improve the generator. Its central contribution is the reward specification; generation remains ordinary autoregressive code prediction.

Refer to caption
Figure 1: Comparison of evaluation metrics and the RULER rubric. Left: Standard scalar metrics (CLIP and Aesthetic) mis-rank or fail to penalize a broken SVG, whereas our Rubric score aligns with human preference. Right: The structured six-item instance-aware rubric spanning semantic, visual, and stylistic axes.

2 Related Work

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.
Reward Paradigm Representative Metric Visual Semantic GT-Free Multi-Axis Instance-
Granularity Granularity Aware
Pixel-based SSIM / PSNR  Wang et al. (2004) Fine-grained N/A ✗ ✗ ✗
Rule-based Code Length  Rodriguez et al. (2025b) N/A N/A ✓ ✗ ✗
Embedding-based CLIPScore  Hessel et al. (2021) Coarse-grained Coarse-grained ✓ ✗ ✗
Rubric-based Universal Rubric  He et al. (2026) Fine-grained Fine-grained ✗ ✓ ✗
Rubric-based Instance-aware Rubric (RULER) Fine-grained Fine-grained ✓ ✓ ✓

2.1 SVG Code Generation

Earlier methods learn vector representations, optimize paths through a diffusion model, or generate SVG programs directly. VectorFusion and SVGDreamer use iterative optimization; IconShop and newer language models generate token sequences. These approaches establish generation capability, while RULER addresses the feedback used to train it [4], [15], [37], [34].

2.2 Rubric-Based Evaluation and Rewards

Rubrics expose separate quality requirements instead of hiding them inside one learned score. Previous work applies this idea to language evaluation and reinforcement learning. RULER adapts the criteria to each visual prompt, where valid solutions vary in shape, composition, and style. Explicit criteria still require testing: optimizing an imperfect reward can exploit its weaknesses [12], [10], [3], [8].

Refer to caption
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).

3 RULER

The pipeline generates and caches a rubric for each training prompt, samples several SVG programs, renders them, and scores each rendering. Group-relative reinforcement learning then increases the probability of better candidates. The rubric generator, training judge, and final evaluation judge have distinct roles.

3.1 Task Definition

Generation is modeled as an MDP: the state is the instruction plus the current code prefix, each action appends a token, and the completed program is rendered for a terminal reward. No externally supplied target image is required.

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

3.2 Human Assessment of Existing Metrics

The authors collect 900 images from three generators on 300 prompts and compare automated metrics with validated human ratings. Rubric scores achieve Spearman correlation 0.7929, versus 0.6051 for aesthetic scores and 0.5518 for CLIP. Pairwise ranking agreement is also higher. This supports the rubric on this sample; it does not establish universal agreement with human taste.

3.3 Instance-Aware Rubric Generation

Claude Opus 4.6 first creates an ideal SVG, then six prompt-specific criteria. The criteria cover concept recognition, essential components, silhouette, composition, execution quality, and stylistic coherence, with weights 5, 5, 5, 4, 5, and 5. They describe design intentions rather than prescribe exact pixels or paths. The internally generated example guides construction and filtering; the training judge does not receive it as a reference image [1].

Equation 1
C𝒬={(ck,wk)}k=16,C_{\mathcal{Q}}=\{(c_{k},w_{k})\}_{k=1}^{6}, (1)

3.4 Policy Optimization

Reward Calculation.

A frozen Qwen3-VL-8B receives the prompt, its rubric, and the rendered candidate. It assigns each item a satisfaction score between zero and one. The normalized weighted sum becomes the candidate’s reward; the default reward adds no separate CLIP, aesthetic, or text-penalty term [2].

Equation 2
ℛ⁡(ℐgen)=∑k=16wk​sk∑k=16wk.\mathcal{R}(\mathcal{I}_{\text{gen}})=\frac{\sum_{k=1}^{6}w_{k}s_{k}}{\sum_{k=1}^{6}w_{k}}. (2)

Group Relative Advantage.

GRPO samples eight candidates per instruction and normalizes their rewards within the group. A clipped policy objective favors candidates with positive relative advantage. The displayed formulation uses a sequence-level importance ratio; the experimental configuration omits KL regularization [27], [11], [26].

Equations 3–4
Ai=ℛi−mean⁡({ℛj}j=1G)std⁡({ℛj}j=1G)+ϵ.A_{i}=\frac{\mathcal{R}_{i}-\mathrm{mean}(\{\mathcal{R}_{j}\}_{j=1}^{G})}{\mathrm{std}(\{\mathcal{R}_{j}\}_{j=1}^{G})+\epsilon}. (3)
ℒGRPO(θ)=𝔼𝒬[1G∑i=1Gmin(ρiAi,clip(ρi,1−ϵ,\displaystyle\mathcal{L}_{\text{GRPO}}(\theta)=\mathbb{E}_{\mathcal{Q}}\biggl[\frac{1}{G}\sum_{i=1}^{G}\min\bigl(\rho_{i}A_{i},\;\text{clip}(\rho_{i},1-\epsilon, (4)
1+ϵ)Ai)],\displaystyle 1+\epsilon)A_{i}\bigr)],

4 Experiments

Experiments test overall generation quality, reward alternatives, rubric components, sensitivity to model choices, and visual failure modes. The main results use one inference run; separate appendix experiments repeat inference with five random seeds.

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.
Method MMSVG-Illustration MMSVG-Icon
CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow Tokens CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow Tokens
Diffusion-Optimized Methods
VectorFusion 0.313 4.930 0.265 0.611 31.4k 0.309 4.624 0.254 0.510 31.4k
SVGDreamer 0.293 5.119 0.266 0.547 124.3k 0.276 4.850 0.253 0.413 124.3k
Foundation LLMs
Qwen3-8B 0.244 4.309 0.233 0.432 0.4k 0.266 4.424 0.234 0.395 0.3k
Qwen3-32B 0.224 4.392 0.215 0.651 0.5k 0.228 4.514 0.218 0.681 0.3k
DeepSeek-V3 0.285 4.653 0.251 0.686 0.5k 0.290 4.610 0.248 0.673 0.2k
Specialist Models
IconShop 0.236 4.480 0.219 0.262 2.6k 0.285 4.541 0.243 0.586 1.3k
OmniSVG-8B 0.215 4.505 0.221 0.288 6.9k 0.272 4.629 0.244 0.565 5.7k
JanusCoder-8B 0.228 4.323 0.227 0.390 1.0k 0.256 4.432 0.234 0.384 1.0k
RULER 0.249 4.859 0.246 0.693 2.0k 0.269 4.873 0.247 0.683 0.7k

4.1 Experimental Setup

Baselines.

Comparisons include iterative VectorFusion and SVGDreamer, general language models, and specialized IconShop, OmniSVG-8B, and JanusCoder-8B. RULER starts from Qwen3-8B; the appendix also trains a 4-billion-parameter version [15], [37], [34], [39], [29], [38].

Benchmarks.

Tests use 150 MMSVG-Illustration prompts and 150 MMSVG-Icon prompts. Training and test prompts are separately generated within the same benchmark distribution, so these results measure generalization to held-out prompts rather than a new domain [39].

Metrics.

The primary score comes from GPT-5-mini applying a shared, universal rubric, separate from the Qwen3-VL-8B training judge. CLIP, aesthetic scores, and HPS provide additional views. Automated scores include only successfully rendered outputs; render success is reported separately. Token counts measure generated SVG code length, not total training or inference computation [14], [35], [20].

4.2 Main Results (RQ1)

Consistent improvements across baselines.

RULER raises the primary rubric score from 0.432 to 0.693 on Illustration and from 0.395 to 0.683 on Icon. The comparison with identical-base-model reward alternatives supports the contribution of the training reward. Improvements are not uniform across every auxiliary image metric.

Competitive performance on auxiliary metrics, top-tier on the rubric score.

RULER’s single-run rubric scores slightly exceed DeepSeek-V3’s 0.686/0.673 and Qwen3-32B’s 0.651/0.681. These numerical differences should not be read as established statistical superiority. Its outputs average 2.0K/0.7K tokens: far shorter than iterative vectorization outputs, but longer than the untrained Qwen3-8B’s 0.4K/0.3K.

Baseline Win Tie Loss Win Rate
Qwen3-8B 122 14 14 89.7%
Qwen3-32B 84 23 43 66.1%
VectorFusion 73 13 64 53.3%
OmniSVG 93 11 46 66.9%
JanusCoder 139 6 5 96.5%
Table 3: Blinded human preference results. Win rates compare RULER against each baseline on 150 MMSVG-Bench prompts and exclude ties.

Human preference confirms the performance gains.

Across 150 prompts per comparison, human win rates excluding ties are 89.7% against Qwen3-8B, 66.1% against Qwen3-32B, 53.3% against VectorFusion, 66.9% against OmniSVG-8B, and 96.5% against JanusCoder-8B. The VectorFusion comparison is close: 73 wins, 64 losses, and 13 ties.

Table 4: Comparison of different RL reward designs on MMSVG benchmarks. C, A and H denote CLIP, Aesthetic and HPS, respectively.
Method MMSVG-Illustration MMSVG-Icon
CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow Tokens CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow Tokens
Zero-Shot 0.244 4.309 0.233 0.432 0.4k 0.266 4.424 0.234 0.395 0.3k
C+A+H RL 0.196 6.697 0.235 0.464 0.7k 0.232 6.210 0.239 0.262 6.3k
Universal Rubric RL 0.244 4.813 0.241 0.660 0.7k 0.266 4.793 0.244 0.591 0.3k
RULER 0.249 4.859 0.246 0.693 2.0k 0.269 4.873 0.247 0.683 0.7k

4.3 Reward Design Analysis (RQ2)

Reward comparisons hold the Qwen3-8B starting policy and optimization settings fixed. They contrast a mixture of conventional metrics, a universal rubric, and the instance-aware rubric.

RULER delivers the strongest balanced gains.

RULER reaches 0.693/0.683, compared with 0.660/0.591 for reinforcement learning with a universal rubric. The gains are 0.033 on Illustration and 0.092 on Icon, supporting prompt-specific criteria in this setting.

Universal rubrics improve steadily but lack instance-level granularity.

A universal rubric already improves substantially over the untuned model, indicating that structured visual feedback helps even without instance-specific descriptions.

Conventional metrics collapse into reward hacking.

Optimizing the CLIP, aesthetic, and HPS mixture raises its aesthetic scores but reduces Icon’s universal rubric score to 0.262 from 0.395. Outputs can accumulate overlapping paths and lose their intended meaning. Higher proxy rewards therefore need not indicate better instruction following.

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

4.4 Ablation Study (RQ3)

Effectiveness of the three rubric axes.

Removing any rubric axis hurts Illustration performance: the primary score falls from 0.693 to 0.580 without visual quality, 0.599 without rendering style, and 0.623 without semantic fidelity. All three axes contribute under this ablation.

Sensitivity to rubric design.

The stricter, structure-focused Rubric-S scores only 0.536 on Illustration and encourages prompt-related text inside drawings despite an explicit text penalty. Rubric wording and calibration can redirect optimization toward shortcuts; adding criteria does not guarantee a better reward.

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 ±\pm standard deviation over five runs.
Method MMSVG-Illustration MMSVG-Icon
CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow
Qwen3-4B 0.240±0.0020.240{\pm}0.002 4.227±0.0344.227{\pm}0.034 0.231±0.0010.231{\pm}0.001 0.372±0.0170.372{\pm}0.017 0.262±0.0020.262{\pm}0.002 4.402±0.0214.402{\pm}0.021 0.232±0.0010.232{\pm}0.001 0.338±0.0170.338{\pm}0.017
RULER-4B 0.244±0.001\mathbf{0.244{\pm}0.001} 4.616±0.020\mathbf{4.616{\pm}0.020} 0.241±0.001\mathbf{0.241{\pm}0.001} 0.561±0.008\mathbf{0.561{\pm}0.008} 0.268±0.001\mathbf{0.268{\pm}0.001} 4.778±0.019\mathbf{4.778{\pm}0.019} 0.244±0.000\mathbf{0.244{\pm}0.000} 0.559±0.006\mathbf{0.559{\pm}0.006}
Qwen3-8B 0.249±0.0030.249{\pm}0.003 4.299±0.0154.299{\pm}0.015 0.234±0.0010.234{\pm}0.001 0.440±0.0110.440{\pm}0.011 0.266±0.0030.266{\pm}0.003 4.434±0.0154.434{\pm}0.015 0.234±0.0010.234{\pm}0.001 0.394±0.0160.394{\pm}0.016
Qwen3-32B 0.265±0.004\mathbf{0.265{\pm}0.004} 4.395±0.0354.395{\pm}0.035 0.241±0.0000.241{\pm}0.000 0.585±0.0050.585{\pm}0.005 0.279±0.002\mathbf{0.279{\pm}0.002} 4.483±0.0224.483{\pm}0.022 0.242±0.0010.242{\pm}0.001 0.566±0.0260.566{\pm}0.026
JanusCoder 0.221±0.0180.221{\pm}0.018 4.357±0.0534.357{\pm}0.053 0.224±0.0060.224{\pm}0.006 0.313±0.1090.313{\pm}0.109 0.249±0.0060.249{\pm}0.006 4.453±0.0314.453{\pm}0.031 0.230±0.0030.230{\pm}0.003 0.324±0.0400.324{\pm}0.040
RULER-8B 0.247±0.0020.247{\pm}0.002 4.872±0.015\mathbf{4.872{\pm}0.015} 0.246±0.000\mathbf{0.246{\pm}0.000} 0.692±0.007\mathbf{0.692{\pm}0.007} 0.269±0.0010.269{\pm}0.001 4.868±0.020\mathbf{4.868{\pm}0.020} 0.247±0.000\mathbf{0.247{\pm}0.000} 0.662±0.016\mathbf{0.662{\pm}0.016}
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 ±\pm standard deviation over five runs.
Method Rubric Generator CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow
Qwen3-8B – 0.266±0.0030.266{\pm}0.003 4.434±0.0154.434{\pm}0.015 0.234±0.0010.234{\pm}0.001 0.394±0.0160.394{\pm}0.016
RULER Claude-Opus-4.6 0.269±0.0010.269{\pm}0.001 4.868±0.020\mathbf{4.868{\pm}0.020} 0.247±0.000\mathbf{0.247{\pm}0.000} 0.662±0.016\mathbf{0.662{\pm}0.016}
GPT-5.5 0.278±0.001\mathbf{0.278{\pm}0.001} 4.708±0.0164.708{\pm}0.016 0.244±0.0000.244{\pm}0.000 0.646±0.0070.646{\pm}0.007
Refer to caption
Figure 5: Qualitative comparison of scalable vector graphics generation between RULER and four baselines.

4.5 Robustness Analysis (RQ4)

Robustness across base models.

Across five inference seeds, the 4-billion-parameter model improves from 0.372 ± 0.017 to 0.561 ± 0.008 on Illustration and from 0.338 ± 0.017 to 0.559 ± 0.006 on Icon. RULER-8B reaches 0.692 ± 0.007 and 0.662 ± 0.016. These are repeated inference runs, not independent training replications.

Robustness across rubric generators.

Replacing Claude Opus 4.6 with GPT-5.5 for rubric generation and retraining Qwen3-8B preserves gains on Icon: 0.646 ± 0.007 versus 0.662 ± 0.016 with Claude-generated rubrics. The same frozen training judge is retained [1], [21].

4.6 Qualitative Analysis (RQ5)

The illustrated examples show clearer subjects, more coherent compositions, and cleaner vector construction after training. They help explain score changes but remain selected qualitative examples, not a measure of failure frequency.

5 Conclusion

Instance-aware visual rubrics provide useful reinforcement-learning rewards for text-to-SVG generation without externally paired target images. The gains depend on how criteria are written and judged; the unsuccessful Rubric-S variant shows that reward hacking remains possible.

Limitations

Rubric generation inherits the frontier model’s preferences, and the frozen visual judge can introduce further biases. Rendering and judging every rollout add training cost. Fixed quality axes may undervalue legitimate artistic choices. Evaluation covers two tasks in one benchmark family; performance across other visual domains and independently repeated training runs remains unestablished.

Acknowledgments

The authors acknowledge support from the Ant Group internship program and feedback from colleagues.

Appendix A Training Details

Training uses VERL and vLLM with thinking disabled, eight rollouts per prompt, sampling temperature 1.0, and maximum prompt/response lengths of 512/4,096 tokens. The batch size is 128, learning rate 10⁻⁶, and entropy coefficient 0.001, with no KL term. Experiments use BF16, gradient checkpointing, and tensor parallelism across eight NVIDIA H800 GPUs with 80 GB each.

Appendix B Dataset Construction

From 32,000 sampled training prompts, rubric construction removes malformed responses and cases where the generated ideal SVG scores below 0.9 under its own rubric. This retains 19,531 Icon and 11,206 Illustration examples, or 30,737 total. The ideal images are generated for this process, not supplied as human ground truth. Filtering couples the dataset to the generator’s and judge’s preferences.

Appendix C Reward Computation and SVG Rendering

CairoSVG 2.9.0 renders candidates at 512 × 512 pixels on a white background. Render failures and effectively blank images receive zero reward. Successful images are judged item by item and aggregated with the displayed weighted mean. Rubric-S’s negative text penalty enters only the numerator; its denominator uses the five positive weights.

Equation
R=∑i=16wi​si∑i=16wi,R=\frac{\sum_{i=1}^{6}w_{i}s_{i}}{\sum_{i=1}^{6}w_{i}},

Appendix D Human Evaluation

Two human studies answer different questions: whether metrics track holistic image ratings, and whether people prefer RULER’s final outputs to competing generators.

D.1 Human Alignment Study

For metric alignment, 10 annotators rate 900 images from Claude Opus 4.6, Qwen3-32B, and Qwen3-8B on 300 prompts. Three validators review unreasonable scores. Each image receives one validated holistic rating from 0 to 100. Correlation and pairwise ranking agreement are computed against those ratings.

D.2 Final-Output Human Preference Study

The preference study compares RULER with five baselines on 150 prompts, producing 750 blinded, randomly ordered pairs distributed among three annotators; each pair is judged once. Judges consider meaning, visual quality, and rendering coherence. Ties are excluded from win rates. Unlike automated quality scores, these comparisons penalize a render failure when the opposing output succeeds.

Equation
WinRate=WinWin+Loss,\mathrm{WinRate}=\frac{\mathrm{Win}}{\mathrm{Win}+\mathrm{Loss}},
Comparison Win Tie Loss Win Rate
RULER vs. Qwen3-8B 122 14 14 89.7%
RULER vs. Qwen3-32B 84 23 43 66.1%
RULER vs. VectorFusion 73 13 64 53.3%
RULER vs. OmniSVG 93 11 46 66.9%
RULER vs. JanusCoder 139 6 5 96.5%
Table 7: Blinded pairwise human preference results on 150 MMSVG-Bench prompts for each baseline comparison. Win rates are computed after excluding ties.

Appendix E Additional Evaluation and Robustness Analyses

E.1 Stability Across Inference Runs

Five independently seeded inference runs report means and standard deviations. They assess sampling variability for trained policies, not variation from retraining the entire method. These repeated-run averages differ from the single-run main table.

E.2 Render Success Rates

RULER renders successfully on 99.3% of Icon prompts and 100% of Illustration prompts, versus 94.7% and 89.3% for Qwen3-8B. This separate accounting matters because failed renders are excluded from automated quality averages.

Method MMSVG-Icon MMSVG-Illustration
RULER 99.3% 100.0%
OmniSVG 99.3% 99.3%
DeepSeek-V3 100.0% 99.3%
JanusCoder 68.0% 68.7%
Qwen3-8B 94.7% 89.3%
Qwen3-32B 99.3% 100.0%
IconShop 100.0% 100.0%
VectorFusion 100.0% 100.0%
SVGDreamer 100.0% 100.0%
Table 8: Render success rates under the single-run benchmark evaluation.

E.3 Policy-Scale Analysis

Training helps both the 4-billion- and 8-billion-parameter base policies under repeated inference. This supports the method at the two tested scales, without establishing scaling behavior beyond them.

E.4 Rubric Generator Analysis

The generator comparison changes which frontier model writes the training rubrics while retaining the Qwen3-8B policy and Qwen3-VL-8B judge. Both rubric sources improve Icon results, although their averages differ.

Table 9: Robustness to different rubric generators on MMSVG-Icon. Results are averaged over five independently seeded inference runs and reported as mean ±\pm standard deviation.
Method Rubric Generator CLIP↑\uparrow Aesth.↑\uparrow HPS↑\uparrow Rubric↑\uparrow
Qwen3-8B – 0.266±0.0030.266{\pm}0.003 4.434±0.0154.434{\pm}0.015 0.234±0.0010.234{\pm}0.001 0.394±0.0160.394{\pm}0.016
Qwen3-32B – 0.279±0.002\mathbf{0.279{\pm}0.002} 4.483±0.0224.483{\pm}0.022 0.242±0.0010.242{\pm}0.001 0.566±0.0260.566{\pm}0.026
JanusCoder – 0.249±0.0060.249{\pm}0.006 4.453±0.0314.453{\pm}0.031 0.230±0.0030.230{\pm}0.003 0.324±0.0400.324{\pm}0.040
RULER Claude-Opus-4.6 0.269±0.0010.269{\pm}0.001 4.868±0.020\mathbf{4.868{\pm}0.020} 0.247±0.000\mathbf{0.247{\pm}0.000} 0.662±0.016\mathbf{0.662{\pm}0.016}
GPT-5.5 0.278±0.0010.278{\pm}0.001 4.708±0.0164.708{\pm}0.016 0.244±0.0000.244{\pm}0.000 0.646±0.0070.646{\pm}0.007

E.5 Comparison with Iterative Diffusion Methods

Iterative diffusion methods generate much longer path programs. The authors report VectorFusion averaging 72.1 minutes per sample in their implementation, while RULER uses one autoregressive generation pass. No corresponding numerical RULER latency is supplied, so the paper does not establish an end-to-end speedup factor.

Appendix F Rubric Design Details

F.1 Ground-Truth-Free Rubric Construction

Rubric construction takes only the instruction as external input but internally synthesizes an ideal SVG before writing criteria. Reward computation later evaluates a candidate against the rubric without comparing it with that ideal. Criteria target recognizable design intentions, leaving room for different valid drawings.

F.2 Rubric-S

Rubric-S emphasizes readability, components, shape complexity, internal details, and layering. A sixth item penalizes text hints with weight −2; a higher score on that item means stronger evidence of a shortcut. Its lower quality score and increased textual shortcuts demonstrate sensitivity to the reward specification.

Appendix G Rubric Generation Cost

The cached training rubrics cost $2,013.12 to generate, averaging $0.0655 per retained prompt. The reported 60.18 million input and 68.49 million output tokens include ideal SVG generation as well as rubric text. This is preprocessing cost, not total training cost: the frontier generator is absent during optimization, but rendering and visual judging continue for each rollout.

Statistic MMSVG-Icon MMSVG-Illustration Overall
Prompts 19,531 11,206 30,737
Avg. Input Tokens 1,959 1,957 1,958
Avg. Rubric Tokens 1,010 1,057 1,027
Avg. Output Tokens 1,934 2,740 2,228
Total Cost $1,135.77 $877.35 $2,013.12
Cost / Prompt $0.0582 $0.0783 $0.0655
Table 10: Rubric-generation cost for the RL training data. Avg. Output Tokens includes both the generated ideal SVG and rubric; Avg. Rubric Tokens counts the rubric portion only.

Appendix H Prompt Templates

H.1 Rubric Generation Prompt

The preserved template requests an ideal SVG followed by six weighted, independently scorable criteria. It emphasizes observable design intentions and calibrated scoring instead of exact reconstruction.

Rubric Generation for Text-to-SVG Evaluation, panel 1 of 3
Rubric Generation for Text-to-SVG Evaluation, panel 2 of 3
Rubric Generation for Text-to-SVG Evaluation, panel 3 of 3
Rubric Generation for Text-to-SVG Evaluation

H.2 Rubric-based Judge Prompt

The item-level judge receives the instruction, candidate rendering, and instance-specific rubric. It returns reasoning and a zero-to-one satisfaction score for each criterion, using only the stated requirements.

Rubric-based VLM Judging
Rubric-based VLM Judging

H.3 Universal Rubric Judge Prompt

The universal evaluation prompt applies shared semantic, visual, and rendering criteria to all candidates and returns an overall score. It also discourages using written labels as substitutes for drawing the requested subject.

Universal Rubric VLM Judging, panel 1 of 3
Universal Rubric VLM Judging, panel 2 of 3
Universal Rubric VLM Judging, panel 3 of 3
Universal Rubric VLM Judging

H.4 Rubric-S Generation Prompt

The alternative generation prompt specifies five structural criteria with stricter scoring guides. The separate preserved panel supplies the fixed negative-weight text-hint penalty used during training.

Rubric-S: Structural Rubric Generation, panel 1 of 2
Rubric-S: Structural Rubric Generation, panel 2 of 2
Rubric-S: Structural Rubric Generation
Fixed Text-Hint Penalty Item
Fixed Text-Hint Penalty Item
Concept demonstration

Open image ↗