pith. sign in

arxiv: 2605.25447 · v1 · pith:HWPBFIVKnew · submitted 2026-05-25 · 💻 cs.CL

GeoSVG-RL: Geometry-Aware Reinforcement Learning for Layout-Constrained Text-to-SVG Diagram Generation

Sifan Li , Yujun Cai , Hongkai Chen , Yiwei Wang This is my paper

Pith reviewed 2026-06-29 22:10 UTC · model grok-4.3

classification 💻 cs.CL
keywords text-to-SVGreinforcement learningdiagram generationgeometric constraintslayout planningSVG generationpolicy optimizationvector graphics
0
0 comments X

The pith

A reinforcement learning method optimizes text-to-SVG generation against browser-rendered geometric rewards to improve diagram reliability.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models often produce SVG diagrams with misaligned connectors, overlapping text, or elements outside the canvas, making them unusable despite fluent code generation. GeoSVG-RL first generates a structured layout plan as a geometric contract, then produces SVG code and evaluates it through a browser-backed verifier that scores six dimensions including rendering validity, anchor placement, text containment, and graph consistency. Group Relative Policy Optimization updates the policy using relative quality among multiple sampled outputs, starting from a supervised warm-start on synthetic data. The approach yields higher local geometric precision and better preservation of graph connectivity than prior systems.

Core claim

Optimizing the generation policy with explicit, executable geometric feedback from rendered SVGs rather than token-level likelihood produces diagrams with substantially improved structural reliability, particularly in arrow-anchor accuracy and text-in-box containment.

What carries the argument

The browser-backed verifier that calculates fine-grained rewards on rendering validity, canvas fitting, precise anchor placement, text containment, graph consistency, and code cleanliness, paired with Group Relative Policy Optimization on multiple candidates per prompt.

If this is right

  • Arrow-anchor accuracy and text containment rates increase substantially after the RL stage.
  • Generated diagrams maintain graph connectivity more reliably than models trained only on token likelihood.
  • The method establishes a pathway from synthetic-data warm-start to production-grade technical illustrations.
  • Multi-candidate sampling with relative ranking enables stable policy updates without absolute reward scales.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The layout-plan-plus-verifier pattern could transfer to other spatially constrained code outputs such as HTML layouts or circuit diagrams.
  • If the verifier dimensions prove composable, the same reward structure might support iterative refinement loops in interactive diagram tools.
  • Larger base models fine-tuned with this geometric signal may close remaining gaps on highly complex multi-object scenes.

Load-bearing premise

The browser-backed verifier supplies accurate and unbiased rewards across the six dimensions that reliably improve the policy without creating new failure modes.

What would settle it

A side-by-side measurement on a held-out prompt set of anchor placement error rates and broken graph connections in SVGs from the RL model versus the supervised baseline.

Figures

Figures reproduced from arXiv: 2605.25447 by Hongkai Chen, Sifan Li, Yiwei Wang, Yujun Cai.

Figure 1
Figure 1. Figure 1: The GeoSVG-RL framework adopts a plan-then-generate approach, where a structured [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the GeoSVG-RL framework. Given a textual prompt [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance gains of GeoSVG-RL over the supervised warm start. Specifically, our method GeoSVG-RL achieves 78.6 AAcc, 0.096 AEE, 83.0 TBR, 3.3 TPVR, and 90.4 E-F1, demonstrating substantial gains in arrow anchoring, text containment, and graph consistency. These results suggest that exe￾cutable rewards effectively optimize the con￾straints essential for diagram usability. Beyond local alignment, GeoSVG-RL … view at source ↗
Figure 4
Figure 4. Figure 4: Examples of text containment and arrow anchoring metrics. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
read the original abstract

Generating structured, editable diagrams remains a significant challenge for contemporary large language models, despite their proficiency in general-purpose vector code generation. The primary difficulty lies in the structural fragility of the output; minor errors such as misaligned connector endpoints, text labels overlapping borders, or complex layouts drifting beyond the canvas boundaries render the resulting SVG files functionally unusable for professional applications. To address these issues, we introduce GeoSVG-RL, a specialized reinforcement learning framework designed for layout-constrained text-to-SVG generation. Unlike standard training objectives that rely solely on maximizing token-level likelihood, our approach optimizes the policy against explicit, executable geometric feedback. The model first produces a structured layout plan that serves as a geometric contract for the subsequent generation of the SVG code. This code is then rendered through a browser-backed verifier, enabling the calculation of fine-grained rewards across six critical dimensions: rendering validity, canvas fitting, precise anchor placement, text containment, graph consistency, and code cleanliness. We utilize Group Relative Policy Optimization (GRPO) to refine the model, sampling multiple candidates per prompt to facilitate updates based on relative quality. Starting from a supervised warm-start phase on synthetic data, GeoSVG-RL achieves substantial gains in structural reliability, particularly in arrow-anchor accuracy and text-in-box rates. Quantitative evaluations demonstrate that our method consistently outperforms current state-of-the-art systems in local geometric precision and the preservation of graph connectivity, providing a robust pathway toward automated yet reliable technical illustration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces GeoSVG-RL, a reinforcement learning framework for text-to-SVG diagram generation that first produces a structured layout plan and then optimizes the policy via Group Relative Policy Optimization (GRPO) against rewards from a browser-backed verifier. The verifier supplies scalar feedback across six dimensions (rendering validity, canvas fitting, precise anchor placement, text containment, graph consistency, and code cleanliness). The work starts from a supervised warm-start on synthetic data and claims substantial gains in structural reliability, with quantitative evaluations showing consistent outperformance over current state-of-the-art systems in local geometric precision and preservation of graph connectivity.

Significance. If the central empirical claims hold, the approach offers a concrete route to more reliable automated technical illustration by replacing token-level likelihood with executable geometric feedback. The use of relative policy optimization over multiple samples per prompt and the explicit multi-dimensional reward design are potentially reusable ideas for other structured generation tasks. However, the significance is limited by the absence of any reported numbers, baselines, or validation of the verifier itself.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'quantitative evaluations demonstrate that our method consistently outperforms current state-of-the-art systems in local geometric precision and the preservation of graph connectivity' is stated without any supporting numbers, tables, baselines, or dataset details. This makes the primary empirical contribution impossible to assess from the manuscript as presented.
  2. [Method] Method description (six reward dimensions): the claim that the browser-backed verifier supplies reliable, unbiased rewards across rendering validity, canvas fitting, precise anchor placement, text containment, graph consistency, and code cleanliness lacks any ablation, error analysis, correlation with human judgment, or external validation. Because GRPO updates are driven directly by these scalar rewards, systematic bias in any dimension (e.g., lenient path parsing or incorrect bounding-box computation) would render the reported gains artifacts of reward hacking rather than genuine structural improvement.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We address each major comment below and commit to revisions that strengthen the presentation of our empirical results and the validation of the reward components.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'quantitative evaluations demonstrate that our method consistently outperforms current state-of-the-art systems in local geometric precision and the preservation of graph connectivity' is stated without any supporting numbers, tables, baselines, or dataset details. This makes the primary empirical contribution impossible to assess from the manuscript as presented.

    Authors: We agree that the abstract would be improved by including concrete numerical support to allow immediate assessment of the claims. The body of the manuscript contains the full quantitative results, including tables with metrics on geometric precision and graph connectivity, comparisons against baselines, and details on the synthetic dataset. To address the concern directly, we will revise the abstract to incorporate key performance numbers and a brief description of the evaluation setup and baselines. revision: yes

  2. Referee: [Method] Method description (six reward dimensions): the claim that the browser-backed verifier supplies reliable, unbiased rewards across rendering validity, canvas fitting, precise anchor placement, text containment, graph consistency, and code cleanliness lacks any ablation, error analysis, correlation with human judgment, or external validation. Because GRPO updates are driven directly by these scalar rewards, systematic bias in any dimension (e.g., lenient path parsing or incorrect bounding-box computation) would render the reported gains artifacts of reward hacking rather than genuine structural improvement.

    Authors: We recognize that explicit validation of the verifier is necessary to substantiate the reliability of the rewards and to address potential concerns about bias or reward hacking. The rewards are computed from executable rendering in a browser environment, which provides objective geometric feedback. Nevertheless, we will add an ablation study on the six reward dimensions, an error analysis of the verifier outputs, and a correlation analysis with human judgments of diagram quality to the revised manuscript. This will include details on how each dimension is computed and any observed limitations. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical RL optimization against external browser verifier.

full rationale

The provided abstract and description present GeoSVG-RL as an RL method (using GRPO) that optimizes against explicit rewards computed by an external browser-backed verifier across six rendering dimensions. No equations, fitted parameters renamed as predictions, self-citations as load-bearing premises, or ansatzes are described. The central claim of outperformance rests on empirical results from this external feedback loop rather than any derivation that reduces to its own inputs by construction. This is the standard case of a self-contained empirical method.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the six reward dimensions and browser verifier are presented as engineering choices without further decomposition.

pith-pipeline@v0.9.1-grok · 5802 in / 1078 out tokens · 24446 ms · 2026-06-29T22:10:20.023503+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 12 canonical work pages

  1. [1]

    Svgenius: Benchmarking llms in svg understanding, editing and generation.arXiv preprint arXiv:2506.03139, 2025

    Siqi Chen, Xinyu Dong, Haolei Xu, Xingyu Wu, Fei Tang, Hang Zhang, Yuchen Yan, Lin- juan Wu, Wenqi Zhang, Guiyang Hou, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Svgenius: Benchmarking llms in svg understanding, editing and generation.arXiv preprint arXiv:2506.03139, 2025

    work page arXiv 2025

  2. [2]

    Weld, and Ranjay Krishna

    Qijia He, Xunmei Liu, Hammaad Memon, Ziang Li, Zixian Ma, Jaemin Cho, Jason Ren, Daniel S. Weld, and Ranjay Krishna. Vfig: Vectorizing complex figures in svg with vision- language models.arXiv preprint arXiv:2603.24575, 2026

    work page arXiv 2026

  3. [3]

    Unisvg: A unified dataset for vector graphic understanding and generation with multimodal large language models

    Jinke Li, Jiarui Yu, Chenxing Wei, Hande Dong, Qiang Lin, Liangjing Yang, Zhicai Wang, and Yanbin Hao. Unisvg: A unified dataset for vector graphic understanding and generation with multimodal large language models. InProceedings of the 33rd ACM International Conference on Multimedia, MM ’25, page 13156–13163, New York, NY , USA, 2025. Association for Com...

    2025

  4. [4]

    Diagrameval: Evaluating llm-generated diagrams via graphs

    Chumeng Liang and Jiaxuan You. Diagrameval: Evaluating llm-generated diagrams via graphs. arXiv preprint arXiv:2510.25761, 2025

    work page arXiv 2025

  5. [5]

    Autofigure-edit: Generating editable scientific illustration.arXiv preprint arXiv:2603.06674, 2026

    Zhen Lin, Qiujie Xie, Minjun Zhu, Shichen Li, Qiyao Sun, Enhao Gu, Yiran Ding, Ke Sun, Fang Guo, Panzhong Lu, Zhiyuan Ning, Yixuan Weng, and Yue Zhang. Autofigure-edit: Generating editable scientific illustration.arXiv preprint arXiv:2603.06674, 2026

    work page arXiv 2026

  6. [6]

    TechING: Towards real world technical image understanding via VLMs

    Tafazzul Nadeem, Bhavik Shangari, Manish Rai, Gagan Raj Gupta, and Ashutosh Modi. TechING: Towards real world technical image understanding via VLMs. In Vera Demberg, Kentaro Inui, and Lluís Marquez, editors,Findings of the Association for Computational Linguistics: EACL 2026, pages 2720–2749, Rabat, Morocco, March 2026. Association for Computational Linguistics

    2026

  7. [7]

    Svgeditbench: A benchmark dataset for quantitative assessment of llm’s svg editing capabilities

    Kunato Nishina and Yusuke Matsui. Svgeditbench: A benchmark dataset for quantitative assessment of llm’s svg editing capabilities. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 8142–8147, June 2024

    2024

  8. [8]

    Svgeditbench v2: A benchmark for instruction-based svg editing, 2025

    Kunato Nishina and Yusuke Matsui. Svgeditbench v2: A benchmark for instruction-based svg editing, 2025

    2025

  9. [9]

    Vectorgym: A multitask benchmark for svg code generation, sketching, and editing.arXiv preprint arXiv:2603.29852, 2026

    Juan Rodriguez, Haotian Zhang, Abhay Puri, Tianyang Zhang, Rishav Pramanik, Meng Lin, Xiaoqing Xie, Marco Terral, Darsh Kaushik, Aly Shariff, Perouz Taslakian, Spandana Gella, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. Vectorgym: A multitask benchmark for svg code generation, sketching, and editing.arXiv preprint arXiv:2603.29852, 2026

    work page arXiv 2026

  10. [10]

    Rodriguez, Abhay Puri, Shubham Agarwal, Issam H

    Juan A. Rodriguez, Abhay Puri, Shubham Agarwal, Issam H. Laradji, Pau Rodriguez, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. Starvector: Generating scalable vector graphics code from images and text.arXiv preprint arXiv:2312.11556, 2023

    work page arXiv 2023

  11. [11]

    Juan A. Rodriguez, Haotian Zhang, Abhay Puri, Rishav Pramanik, Aarash Feizi, Pascal Wich- mann, Arnab Kumar Mondal, Mohammad Reza Samsami, Rabiul Awal, Perouz Taslakian, Span- dana Gella, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. Rendering- aware reinforcement learning for vector graphics generation. InThe Thirty-ninth Annual Conf...

    2026

  12. [12]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024

    2024

  13. [13]

    Introsvg: Learning from rendering feedback for text-to-svg generation via an introspective generator-critic framework.arXiv preprint arXiv:2603.09312, 2026

    Feiyu Wang, Jiayuan Yang, Zhiyuan Zhao, Da Zhang, Bingyu Li, Peng Liu, and Junyu Gao. Introsvg: Learning from rendering feedback for text-to-svg generation via an introspective generator-critic framework.arXiv preprint arXiv:2603.09312, 2026

    work page arXiv 2026

  14. [14]

    Svgen: Interpretable vector graphics generation with large language models.arXiv preprint arXiv:2508.09168, 2025

    Feiyu Wang, Zhiyuan Zhao, Yuandong Liu, Da Zhang, Junyu Gao, Hao Sun, and Xuelong Li. Svgen: Interpretable vector graphics generation with large language models.arXiv preprint arXiv:2508.09168, 2025. 10

    work page arXiv 2025

  15. [15]

    Reliable reasoning in svg-llms via multi-task multi-reward reinforcement learning.arXiv preprint arXiv:2603.16189, 2026

    Haomin Wang, Qi Wei, Qianli Ma, Shengyuan Ding, Jinhui Yin, Kai Chen, and Hongjie Zhang. Reliable reasoning in svg-llms via multi-task multi-reward reinforcement learning.arXiv preprint arXiv:2603.16189, 2026

    work page arXiv 2026

  16. [16]

    Empowering llms to understand and generate complex vector graphics.arXiv preprint arXiv:2412.11102, 2024

    Ximing Xing, Juncheng Hu, Guotao Liang, Jing Zhang, Dong Xu, and Qian Yu. Empowering llms to understand and generate complex vector graphics.arXiv preprint arXiv:2412.11102, 2024

    work page arXiv 2024

  17. [17]

    Reason-svg: Enhancing structured reasoning for vector graphics generation with reinforcement learning, 2026

    Ximing Xing, Ziteng Xue, Yandong Guan, Jing Zhang, Dong Xu, and Qian Yu. Reason-svg: Enhancing structured reasoning for vector graphics generation with reinforcement learning, 2026

    2026

  18. [18]

    Omnisvg: A unified scalable vector graphics generation model.arXiv preprint arXiv:2504.06263, 2025

    Yiying Yang, Wei Cheng, Sijin Chen, Xianfang Zeng, Fukun Yin, Jiaxu Zhang, Liao Wang, Gang Yu, Xingjun Ma, and Yu-Gang Jiang. Omnisvg: A unified scalable vector graphics generation model.arXiv preprint arXiv:2504.06263, 2025

    work page arXiv 2025

  19. [19]

    Structural evaluation metrics for svg generation via leave-one-out analysis, 2026

    Haonan Zhu, Adrienne Deganutti, Elad Hirsch, and Purvanshi Mehta. Structural evaluation metrics for svg generation via leave-one-out analysis, 2026

    2026

  20. [20]

    Autofigure: Generating and refining publication-ready scientific illustrations

    Minjun Zhu, Zhen Lin, Yixuan Weng, Panzhong Lu, Qiujie Xie, Yifan Wei, Sifan Liu, Qiyao Sun, and Yue Zhang. Autofigure: Generating and refining publication-ready scientific illustrations. arXiv preprint arXiv:2602.03828, 2026

    work page arXiv 2026

  21. [21]

    Svgauge: Towards human-aligned evaluation for svg generation, 2025

    Leonardo Zini, Elia Frigieri, Sebastiano Aloscari, Marcello Generali, Lorenzo Dodi, Robert Dosen, and Lorenzo Baraldi. Svgauge: Towards human-aligned evaluation for svg generation, 2025. 11 A Implementation Details A.1 Base Model The SVG generator is initialized from Qwen2.5-Coder-7B-Instruct, a pretrained autoregressive code model with 7B parameters. Thi...

    2025