VectorGym

01 · Tasks

Four tasks for visual code generation

An SVG is a program: its code specifies geometry, style and structure, and renders directly to an image. One benchmark can therefore assess visual fidelity, syntactic validity and structured manipulation together.

02 · Data

Real-world SVGs annotated by human experts

VectorGym draws on SVG-Stack, a collection of SVGs from GitHub that spans icons, diagrams, emojis, fonts, logotypes and complex illustrations and keeps their original primitives. More than 20 annotators from two vendors, with expertise in design, vector graphics and coding, created the sketches, edits and descriptions.

Candidate SVGs

7,000

from the SVG-Stack training split

Training set

6.5k

after filtering; 100 of these form the validation set

Test

300

from the SVG-Stack test split

Annotators

20+

design, vector graphics and coding

Filters retained SVGs of 2k to 8k tokens with normalized color entropy above 0.55; a random subsample was then inspected by hand.

VG-Edit · test examples

03 · Method

Multi-task reinforcement learning with rendering and text-similarity rewards

The baseline fine-tunes Qwen3-VL 8B Instruct on all four tasks jointly with Reinforcement Learning from Rendering Feedback (RLRF), using the VectorGym training split.

1. Reward

   Rewards from rendered output

   For the three generation tasks, predicted and reference SVGs are rendered and compared with perceptual similarity and pixel-space distances. For captioning, the reward is the BGE-M3 embedding similarity between predicted and reference captions.

2. Optimizer

   GRPO with batch-level normalization

   Standard GRPO normalizes rewards within each group of sampled responses. This variant divides the centered rewards by the standard deviation over the whole mini-batch:

   \[A^{\text{batch}}_{i}=\frac{r_{i}-\operatorname{mean}\big(\{r_{j}\}_{j=1}^{K}\big)}{\operatorname{std}\big(\{r_{j}\}_{j=1}^{N\times K}\big)}\]

   \(K\) is the number of sampled responses per prompt and \(N\times K\) the number of samples in the mini-batch.

3. Curriculum

   Length-based curriculum

   Response length is used as a proxy for difficulty. Samples are sorted by length within each task, and mini-batches draw tasks in proportion to their frequency, so training progresses from short to long examples with the tasks kept balanced.

Rollouts per step

1,344

168 samples × 8 rollouts

Training

600 iters

8 × H200, about two days (384 GPU-hours)

VectorGym score

58.74 → 66.05

+7.31 over the base 8B model; inference cost unchanged

04 · Results

Large gaps remain across tasks; the RL baseline matches GPT-4o

The VectorGym score is the mean of four task scores (0 to 100). Gemini 3 Pro ranks first (73.17). The RL-trained Qwen3-VL 8B reaches 66.05, above GPT-4o (64.93) and the much larger Qwen3-VL 235B Instruct (62.32).

VLM judge vs. human ratings

05 · Citation

Cite VectorGym

@misc{rodriguez2026vectorgym,
  title         = {VectorGym: A Multi-Task Benchmark for SVG Code Generation,
                   Sketching and Editing},
  author        = {Joan Rodriguez and Haotian Zhang and Abhay Puri and Haoran Dai and
                   Tianyang Zhang and Meng Lin and Rishav Pramanik and Xiaoqing Xie and
                   Marco Terral Rodriguez and Darsh Kaushik and Aly Shariff and
                   Perouz Taslakian and Spandana Gella and Sai Rajeswar and
                   David Vazquez and Christopher Pal and Marco Pedersoli},
  year          = {2026},
  eprint        = {2603.29852},
  archivePrefix = {arXiv},
  primaryClass  = {cs.GR},
  note          = {Accepted to EMNLP 2026}
}