VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing

# VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing
Source: [https://arxiv.org/html/2605.15677](https://arxiv.org/html/2605.15677)
Peijie DongZhenheng TangSong TangYuyao ZhaiKaitao LinLiang ChenYuhang GaiYuyu LuoQiang WangXiaowen Chu

###### Abstract

Despite the rapid advancements in Vision\-Language Models \(VLMs\), a critical gap remains in their ability to handle structured, controllable diagrammatic tasks essential for professional workflows\. Existing methods predominantly rely on pixel\-based synthesis, which operates in probabilistic pixel spaces and is inherently limited in editability and fidelity\. Instead, we propose a new “Diagram\-as\-Code” paradigm with symbolic logic that leveragesmxGraphExtensible Markup Language \(XML\) for precise diagram generation and editing\. We presentVCG\-Bench, a unified benchmark for visual\-centricmxGraphtasks\. VCG\-Bench comprises: \(1\) a taxonomized dataset of 1,449 diverse diagrams spanning 6 domains and 15 sub\-domains, \(2\) a paradigm definition that integrates Generation \(Vision\-to\-Code\) and Editability \(Code\-to\-Code\), \(3\) a Tailored Evaluation Protocol employing multi\-dimensional metrics such asmxGraphExecution Success Rate, Style Consistency Score \(SCS\), etc\. Experimental results highlight the challenges faced by current State\-of\-the\-Art \(SOTA\) VLMs in structured fidelity and instruction compliance, reflecting their vision and reasoning capabilities\.

Machine Learning, ICML

## 1Introduction

Vision\-Language Models \(VLMs\)\(OpenAI,[2023](https://arxiv.org/html/2605.15677#bib.bib2); Google DeepMind,[2025](https://arxiv.org/html/2605.15677#bib.bib48); Wuet al\.,[2025a](https://arxiv.org/html/2605.15677#bib.bib4); Chenet al\.,[2025b](https://arxiv.org/html/2605.15677#bib.bib70); Liuet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib71); Wanget al\.,[2024a](https://arxiv.org/html/2605.15677#bib.bib72); Baiet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib80)\)have demonstrated impressive capabilities in general multimodal tasks, spanning visual question answering\(Antolet al\.,[2015](https://arxiv.org/html/2605.15677#bib.bib1)\)to open\-ended image generation\(Rombachet al\.,[2022](https://arxiv.org/html/2605.15677#bib.bib22); Podellet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib73)\)\. Through high\-capacity architectures and massive\-scale multimodal pre\-training, modern VLMs have achieved near\-human performance in scene description, optical character recognition \(OCR\), and cross\-modal reasoning\. Yet, moving beyond natural images to abstract, symbolic, and structured visual information remains a key frontier\. This gap is particularly evident for diagrams such as flowcharts, system architectures, scientific figures, and UI mockups—which are pervasive in professional settings\(Panet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib5)\)\. Unlike natural images, diagrams are designed to encode precise semantics through composition, layout, and discrete symbols, requiring models to capture structure rather than appearance\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/symbol_logic.png)Figure 1:Comparison of diagrammatic tasks\.VCG utilizes symbolicmxGraphXML for precise generation and editing, overcoming structural drift in pixel\-based models\.Table 1:Comparative analysis of visual representation formats\.mxGraphbalances semantic richness with structured editability\.![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/framework_official.png)Figure 2:Overview of the VCG\-Bench framework\.Unlike fragmented approaches \(top\), VCG\-Bench \(bottom\) unifies Vision\-to\-Code Generation \(Task 1\) and Instruction\-to\-Patch Editing \(Task 2\)\. Utilizing symbolicmxGraphXML enables precise, low\-cost modifications for professional workflows\.We proposeVCG\-Bench, a unified benchmark designed to evaluate the end\-to\-end capability of VLMs in generating and editingmxGraphXML diagrams\. VCG\-Bench follows a “Data\-Task\-Evaluation” framework\. \(1\) We construct a diverse dataset categorized into 6 major domains \(Academic, Software, Business, etc\.\), covering complex, logic\-intensive layouts\. \(2\) We unify diagrammatic tasks into two complementary streams:Generation\(creating validmxGraphXML from standard images\) andEditability\(modifying existing structures based on instructions\), with anIncremental Modification Strategyto enhance efficiency\. \(3\) We establish a rigorous evaluation protocol beyond simple text matching, incorporating execution success rates, visual style consistency \(SCS\), and semantic fidelity checks via QA\. Our contributions are summarized below:

- •We introduce VCG\-Bench, the first benchmark bridging the gap between vision and editable structured generation\.
- •We curate a high\-quality, taxonomized dataset of 1,449 diagrams with provenance and structural diversity\.
- •We propose a comprehensive evaluation protocol with novel metrics for executability, style consistency, and instruction following, benchmarking SOTA VLMs on diagrammatic tasks\.

To motivate this benchmark design, we first contrast two diagrammatic paradigms\. For diagrammatic tasks, one of the main paradigms is“Diagram\-as\-Pixels”\. It relies on rasterized pixel generation and editing pipelines \(e\.g\., prompt \+ pixel\-editing models such as GPT\-Image and diffusion models\)\. Although it can produce visually plausible results, it is inherentlystochastic, unstructured, and hallucination\-prone, because it operates on pixels rather than symbolic structure\. Fig\.[1](https://arxiv.org/html/2605.15677#S1.F1)illustrates that this often leads to*blurry lines, text distortion, scale/alignment errors,*and even*hallucinated elements*\(e\.g\., inserting an unintended “CDN”\)\. These issues conflict with practical diagram editing workflows that require: \(1\)vector\-accurate outputs\(e\.g\.,SVGormxGraph\), \(2\)low\-cost incremental modifications, and \(3\)frequent iterative editswithout degradation\.

To address these limitations, we adopt a“Diagram\-as\-Code”paradigm, which is morecontrollable, structured, and deterministic\. Instead of editing pixels, we useVLMs\(e\.g\., Gemini\(Google DeepMind,[2025](https://arxiv.org/html/2605.15677#bib.bib48)\), Claude\(Anthropic,[2025](https://arxiv.org/html/2605.15677#bib.bib49)\), and GPT\(OpenAI,[2025](https://arxiv.org/html/2605.15677#bib.bib50)\)\) to convert raster figures into structured representations\. Guided by user instructions, the system performs edits viaVisual Controllable Generation \(VCG\)withsymbolic operations, enabling effective collaboration between LLMs and human users while preserving precise, editable diagram structure—producingclean, high\-fidelityoutput figures \(Fig\.[1](https://arxiv.org/html/2605.15677#S1.F1)\)\. Notably, theDiagram\-as\-Codeparadigm can be integrated with pixel\-level generators to produce more aesthetically pleasing, stylized renderings\.

For“Diagram\-as\-Code”, Tab\.[1](https://arxiv.org/html/2605.15677#S1.T1)compares dominant representation formats \(mxGraph,SVG,HTML/CSS,PPT\)\.SVGis largely geometry\-centric with limited semantics,HTML/CSSoffers only moderate structural control, andPPTis GUI\-centric with restricted programmatic editability\. By contrast,mxGraphmodels diagrams as semantic entities \(nodes/edges\), providinghigh semantic depth,structured editability, andhigh AI controllabilitythrough an interpretable open XML representation, making it well\-suited for AI\-driven reasoning and precise modification\. \(See App\.[D\.2](https://arxiv.org/html/2605.15677#A4.SS2)for further discussion\.\)

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/data_pipeline.png)Figure 3:Overview of the VCG\-Bench data generation framework\. Subfigure \(a\) illustrates the end\-to\-end pipeline for Task 1, from raw web scraping to structured XML\-based rendering\. Subfigure \(b\) details the Task 2 pipeline, which focuses on generating editing\-based reasoning tasks derived from Task 1\.Fig\.[3](https://arxiv.org/html/2605.15677#S1.F3)details the data generation pipeline for VCG\-Bench and shows the systematic process from data collection to structured evaluation\. App\.[B](https://arxiv.org/html/2605.15677#A2)provides detailed descriptions of the data synthesis pipeline, including Stage 1 \(Image\-to\-mxGraphXML\) and Stage 2 \(Instruction Synthesis\)\.

## 2Related Work

Our work advances the domain of multimodal code generation by focusing on themxGraphformat\. We situate it in four strands of related literature: chart and scientific plot generation, UI and symbolic graphics, instruction\-driven image editing, and evaluation methodologies\.

Chart and Scientific Plot Generation\.Diagrammatic reasoning has been advanced by VQA benchmarks \(e\.g\.,AI2D\(Kembhaviet al\.,[2016](https://arxiv.org/html/2605.15677#bib.bib6)\),ChartQA\(Masryet al\.,[2022](https://arxiv.org/html/2605.15677#bib.bib8)\),FigureQA\(Kahouet al\.,[2018](https://arxiv.org/html/2605.15677#bib.bib74)\),MMBench\(Liuet al\.,[2024b](https://arxiv.org/html/2605.15677#bib.bib9)\)\) and chart\-understanding evaluations \(CharXiv\(Wanget al\.,[2024b](https://arxiv.org/html/2605.15677#bib.bib11)\),MMSCI\(Liet al\.,[2024c](https://arxiv.org/html/2605.15677#bib.bib12)\),AIBench\(Liaoet al\.,[2026](https://arxiv.org/html/2605.15677#bib.bib79)\)\)\. Recent work on executable code includesChartMimic\(Yanget al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib14); Niuet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib13)\),Plot2Code\(Wuet al\.,[2025b](https://arxiv.org/html/2605.15677#bib.bib15)\),PlotCraft\(Zhanget al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib16)\),MMCode\(Liet al\.,[2024b](https://arxiv.org/html/2605.15677#bib.bib18)\),Flow2Code\(Heet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib17)\), andRealChart2Code\(Zhanget al\.,[2026](https://arxiv.org/html/2605.15677#bib.bib78)\); most of this line remains centered on rasterized, data\-driven visualizations\. Interactive visualization systems such asHAIChart\(Xieet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib62)\)study human\-AI paired chart creation\. Related to scientific diagrams,Draw with Thought\(Cuiet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib59)\)reconstructs static scientific diagrams into editablemxGraphXML\. VCG\-Bench provides a multi\-domainDraw\.io/mxGraphbenchmark that jointly covers image\-to\-XML generation and instruction\-based XML editing with explicit spatial and topological evaluation\. Tab\.[2](https://arxiv.org/html/2605.15677#S2.T2)positions VCG\-Bench against these benchmarks\.

Table 2:Comparison of VCG\-Bench with related benchmarks\.Edit\.: support for instruction\-based editing\.Fine\-grained: multi\-dimensional evaluation beyond single pass/fail or similarity\.✓\\checkmark: presence;×\\times: absence\.VCG\-Bench is the onlymulti\-domainbenchmark that jointly supportsEdit\.andFine\-grainedevaluation \(Tab\.[2](https://arxiv.org/html/2605.15677#S2.T2)\)\. Prior visual\-code benchmarks focus on single\-domain generation \(execute\-pass or similarity\); PlotCraft\(Zhanget al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib16)\)supports instruction\-based editing and fine\-grained evaluation but remains within data visualization\. We unify multi\-domain diagram coverage with Task 1 \(image\-to\-mxGraphXML\) and Task 2 \(instruction\-based editing\), evaluated via XQA, XDRFR, ESR, and SCS; this design supports both rigorous evaluation and scalable generation of verified visual\-code pairs for model training\.

User Interface and Symbolic Graphics Generation\.UI\-to\-code has evolved from early CNN/RNN systems \(Pix2Code\(Beltramelli,[2018](https://arxiv.org/html/2605.15677#bib.bib20)\),Sketch2Code\(Jainet al\.,[2019](https://arxiv.org/html/2605.15677#bib.bib21)\)\) to diffusion\-based generation\(Rombachet al\.,[2022](https://arxiv.org/html/2605.15677#bib.bib22); Zhanget al\.,[2023b](https://arxiv.org/html/2605.15677#bib.bib23)\)and benchmarks such asDesign2Code\(Siet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib29)\); code LLMs\(Chenet al\.,[2021](https://arxiv.org/html/2605.15677#bib.bib24); Liet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib25); Lozhkovet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib75); Roziereet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib27); Guoet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib28)\)have advanced program synthesis\. In vector graphics,VCode\(Linet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib30)\)andSVGenius\(Chenet al\.,[2025a](https://arxiv.org/html/2605.15677#bib.bib37)\)targetSVG; learned \(DeepSVG\(Carlieret al\.,[2020](https://arxiv.org/html/2605.15677#bib.bib32)\),Im2Vec\(Reddyet al\.,[2021](https://arxiv.org/html/2605.15677#bib.bib33)\)\) and LLM\-based\(Jainet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib34); Vinkeret al\.,[2022](https://arxiv.org/html/2605.15677#bib.bib35); Wuet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib36)\)methods addressSVG\. Recent benchmarks further evaluate SVG editing\(Nishina and Matsui,[2024](https://arxiv.org/html/2605.15677#bib.bib60),[2025](https://arxiv.org/html/2605.15677#bib.bib61)\)and text\-to\-diagram generation/editing\(Weiet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib57)\)\. These adjacent settings focus on SVG/vector graphics or general structured visuals; we target editablemxGraphXML with spatial and syntactic constraints, which preserves higher\-level diagram topology and edit semantics for precise component manipulation\.

Instruction\-Driven Image Editing\.Diffusion\-based editing \(Prompt\-to\-Prompt\(Hertzet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib38)\),InstructPix2Pix\(Brookset al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib39)\),MagicBrush\(Zhanget al\.,[2023a](https://arxiv.org/html/2605.15677#bib.bib40)\)\) has achieved significant progress but suffers from semantic entanglement \(difficulty preserving unedited regions\) and insufficient precision \(character distortion, topological hallucination\)\. We reframe editing as code refactoring onmxGraphXML, preserving exact fidelity outside edited regions and enabling edits that are infeasible in pixel space\. This code\-mediated paradigm thus avoids the semantic entanglement and precision limits of pixel\-level editing while supporting deterministic, localized modifications\.

Evaluation Methodologies for Instruction Following\.Robust evaluation remains a critical challenge; traditional metrics often fail to capture the nuance of complex instructions\.InFoBench\(Qinet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib46); Zhouet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib77)\)introduced the Decomposed Requirements Following Ratio \(DRFR\); multimodal benchmarks\(Yuet al\.,[2023](https://arxiv.org/html/2605.15677#bib.bib41); Yueet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib42),[2025](https://arxiv.org/html/2605.15677#bib.bib43); Liet al\.,[2024a](https://arxiv.org/html/2605.15677#bib.bib44); Xuet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib45); Chenet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib76)\)and domain\-specific work\(Yanget al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib14); Wuet al\.,[2025b](https://arxiv.org/html/2605.15677#bib.bib15); Siet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib29); Zhuoet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib58); Xieet al\.,[2026](https://arxiv.org/html/2605.15677#bib.bib65)\)use fine\-grained criteria, including structured\-visual factuality checks and visualization\-quality assessment\. Drawing inspiration from these, we adopt a decomposed strategy \(XDRFR and related metrics\) for structural and visual fidelity ofDraw\.iooutputs, allowing fine\-grained verification of instruction compliance directly on the generated XML\. Evaluating on the XML structure rather than rendered images avoids visual ambiguities and supports reproducible, attribute\-level checks\. A manual audit of XDRFR \(App\.[A](https://arxiv.org/html/2605.15677#A1)\) confirms high agreement between automated scores and human judgment\.

## 3VCG\-Bench Construction

### 3\.1Data Provenance & Licensing

Our dataset is curated from four sources: Open\-source Template Repositories \(40%\), Open Access Academic Papers \(30%\), Anonymized Corporate Diagrams \(20%\), and Permissively Licensed Web Diagrams \(10%\)\. We incorporate theSALT\-NLP/Design2Codedataset\(Siet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib29)\)from Hugging Face Hub under its original licensing terms\. The dataset, comprisingmxGraphXML source files, rendered high\-resolution images, captions, and editing instructions, is released underCC BY 4\.0\. The evaluation toolkit is released underMIT License\. Web\-crawled data undergoes PII redaction to remove names, IP addresses, and sensitive corporate identifiers\. App\.[B](https://arxiv.org/html/2605.15677#A2)provides details on data sources, licensing policies, and diversity considerations\.

### 3\.2Quality Assurance Pipeline

We employ a “Machine\-Generation, Human\-Verification” pipeline: \(1\)Acquisition & Captioning: Gemini\-3\-Pro generates structured JSON captions of visual elements and layout; \(2\)mxGraphXML Synthesis: during data construction, candidatemxGraphXML files are synthesized from image\-caption pairs by multiple VLM candidates and retained only after downstream quality filters; \(3\)Parser Validation: rule\-based validation against themxGraphschema filters errors; \(4\)Rendering Verification: similarity between renderedIg​e​nI\_\{gen\}and sourceIs​r​cI\_\{src\}is computed viaSigLIP2\(Tschannenet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib47)\)embeddings, discarding samples with cosine similarity<0\.85<0\.85; \(5\)Human Review: annotators verify structural completeness, text accuracy \(\>95%\>95\\%OCR match\), and editability\. This data\-construction process is separate from the formal benchmark evaluation in Sec\.[6](https://arxiv.org/html/2605.15677#S6), where all evaluated models are run on the released Task 1/Task 2 inputs\. From 5,000\+ candidates, parsing and rendering success rates were 60% and 55%, respectively\. After human filtering \(45% acceptance\), 1,449 high\-quality samples were retained, requiring 2\.3 edits per sample on average for alignment\.

Table 3:Dataset Taxonomy\.Summary of the 6 major domains and 15 sub\-domains\.![Refer to caption](https://arxiv.org/html/2605.15677v1/x1.png)Figure 4:Left:Distribution across 15 sub\-domains\.Stratified by difficulty, reflecting structural complexity and element density\. Right:Dataset composition\.
### 3\.3Dataset Taxonomy

The benchmark categorizes 1,449 examples into 6 major domains \(L1\) and 15 sub\-domains \(L2\), ensuring broad coverage of diagrammatic complexity\. App\.[B\.1](https://arxiv.org/html/2605.15677#A2.SS1)provides a detailed breakdown, with summary in Tab\.[3](https://arxiv.org/html/2605.15677#S3.T3); Fig\.[4](https://arxiv.org/html/2605.15677#S3.F4)visualizes distribution across sub\-domains and difficulty\.

## 4Task Definition

We define two core tasks under a unified framework formalized byexecutability constraints: both require model outputs to be validmxGraphXML that can be parsed and rendered\. The following subsections formalize Generation \(Vision\-to\-Code\) and Editability \(Code\-to\-Code\)\.

### 4\.1Generation \(Vision\-to\-Code\)

Definition: Given a raster imageI∈ℝH×W×3I\\in\\mathbb\{R\}^\{H\\times W\\times 3\}and a structured captionSSof visual elements and layout, the model generates a validmxGraphXMLstringCCsuch thatR​e​n​d​e​r​\(C\)Render\(C\)approximatesIIin structure and semantics\.Optimization Goal: The model maximizes likelihoodpθ​\(C\|I,S\)p\_\{\\theta\}\(C\|I,S\)\.Constraints: ①Syntactic Validity:CCmust be valid XML parsable by themxGraphlibrary\. ②Visual Fidelity:R​e​n​d​e​r​\(C\)Render\(C\)must matchIIin topological structure \(node connections\), spatial layout \(bounding boxes\), and textual content\.Output: A rawmxGraphXML string containing ‘mxGraphModel‘ definitions\. App\.[D](https://arxiv.org/html/2605.15677#A4)provides themxGraphXML schema and format details\. Our main Task\-1 setting uses image plus structured caption as input; Sec\.[6\.2](https://arxiv.org/html/2605.15677#S6.SS2)reports an image\-only ablation showing that raw\-image\-to\-mxGraphis substantially harder\.

### 4\.2Editability \(Code\-to\-Code\)

Definition: Given sourcemxGraphXMLCs​r​cC\_\{src\}, renderingIs​r​cI\_\{src\}, and instructionTT\(e\.g\., “Change the decision node from red to blue”\), the model predicts aDifferential PatchPP, yieldingCt​g​t=A​p​p​l​y​\(Cs​r​c,P\)C\_\{tgt\}=Apply\(C\_\{src\},P\)\.Incremental Modification Strategy: Rather than regenerating the fullmxGraphXML, which is token\-expensive and prone to structural drift, we predict only themxGraphXML fragment pairs to be changed\.Patch Schema: We use a deterministicJSON\-based replacement format, where each change contains original and modifiedmxGraphXML fragments:

1\{

2"changes":\[\{

3"original\_fragment":"<mxCellid=\\"s2\_r0\\"\.\.\.

4fillColor=\#FF0000\.\.\.\>\.\.\.</mxCell\>",

5"modified\_fragment":"<mxCellid=\\"s2\_r0\\"\.\.\.

6fillColor=\#0000FF\.\.\.\>\.\.\.</mxCell\>"

7\}\]

8\}

Patch Applier: A deterministic applierΦ​\(C,P\)\\Phi\(C,P\)replaces the matched source fragment inCCwith the target fragment; deletion uses an empty target\. If exact matching fails, fuzzy regex matching ignores whitespace\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/task1_facet_detailed_models.png)Figure 5:Performance scaling and robustness across diagrammatic complexity tiers\. Each panel illustrates the CodeXQA accuracy of a specific model family as task difficulty increases from Easy to Hard\.
### 4\.3Instruction Taxonomy

We develop an instruction generation taxonomy grounded inatomic operations\. We define a set of14 atomic operation categoriesthat encompass diagram editing tasks, organized into four domains: \(1\)Node Attribute Modification\(4 types\): altering color, shape, size, and text; \(2\)Node Structure Operations\(3 types\): insertion, deletion, and relocation; \(3\)Edge Attribute Modification\(3 types\): adjusting color, style, and arrow endpoints; and \(4\)Edge Structure Operations\(4 types\): insertion, deletion, redirection, and path updates\.

Complexity Stratification\.We stratify editing instructions into three tiers by the number of atomic operations required:Easy\(1–2 operations\),Medium\(3–4 operations\), andHard\(5–7 operations\)\. This hierarchical design ensures a robust evaluation across varying modification intensity\. For each sample, we generate instructions for all three tiers\. Instruction synthesis is automated via LLMs, conditioning on the rendered images and themxGraphXML structure\. To mirror real\-world usage, instructions employ natural language descriptors \(e\.g\., referencing node text\) rather than technical identifiers\. App\.[E](https://arxiv.org/html/2605.15677#A5)provides prompt templates for instruction generation, XML editing, and evaluation\.

## 5Evaluation Protocol

We establish an evaluation protocol measuring three dimensions: \(1\)Executability, assessing syntactic validity and renderability; \(2\)Visual Fidelity, verifying structural and semantic reconstruction; and \(3\)Semantic Compliance, evaluating instruction adherence and logic preservation\. All model\-based evaluations employGemini\-3\-Profor consistency\. Among our metrics, SCS is the most exposed to judge\-family effects because it relies on visual judgment, whereas ESR is programmatic and CodeXQA/XDRFR operate over structured XML\-based QA or verification\. We therefore report a cross\-judge validation for SCS in Tab\.[5](https://arxiv.org/html/2605.15677#S5.T5)\. Metrics are summarized in Tab\.[4](https://arxiv.org/html/2605.15677#S5.T4)\. App\.[F](https://arxiv.org/html/2605.15677#A6)provides detailed mathematical formulations, question set construction procedures, and validation protocols\.

Table 4:Metrics Overview\.Lists all evaluation metrics used in VCG\-Bench along with their definitions and inputs\.##### Executability Metric\.

Execution Success Rate \(ESR\)\.We check whether generated code is syntactically valid and renderable\. For codeCC, ESR indicates successful parsing by themxGraphengine and rendering into imageIg​e​nI\_\{gen\}:

ESR=𝕀​\(Parsable​\(C\)∧Renderable​\(C\)\)\\text\{ESR\}=\\mathbb\{I\}\(\\text\{Parsable\}\(C\)\\land\\text\{Renderable\}\(C\)\)\(1\)where𝕀​\(⋅\)\\mathbb\{I\}\(\\cdot\)is the indicator function\. Samples withESR=0\\text\{ESR\}=0receive zero on subsequent metrics, serving as a filter\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/demo2.png)Figure 6:Task 1 qualitative examples\.This figure showcases generated diagrams across multiple domains with their corresponding Style Consistency Scores \(SCS\), SCS rankings, and expert rankings\. The examples demonstrate that the SCS metric rankings and human expert rankings are highly consistent, proving that the SCS metric accurately reflects human aesthetic standards\.
##### Visual Fidelity Metrics\.

Style Consistency Score \(SCS\)\.SCS uses a VLM\-based evaluator \(Gemini\-3\-Pro\) to assess visual alignment, addressing limitations of pixel\-level metrics in capturing structural aesthetics\. ForTask 1 \(Generation\), SCS evaluates three dimensions on a 10\-point scale: \(1\)Visual Style\(color, stroke, shape\); \(2\)Layout Consistency\(spatial topology\); \(3\)Aesthetic Quality\(alignment, balance\)\. Each score is prompted, averaged, and normalized to\[0,1\]\[0,1\]\. ForTask 2 \(Editing\), SCS focuses onStyle ConsistencyandAesthetic Qualityrelative to pre\-edit, ensuring modifications preserve professional standards\.

Robustness of SCS to judge choice\.Since SCS is a judge\-based visual metric, we additionally test whether its relative conclusions are stable across evaluator families\. We randomly sample 100 Task\-1 examples, fix the outputs from four representative generators \(GPT\-5\.2, Gemini\-3\-Pro, Claude\-4\.5\-Sonnet, and GLM\-4\.6V\), and re\-score the same outputs with Gemini, GPT, and Claude judges\. We retain the shared valid subset where all judges return directly comparable scores \(n=78n=78\)\.

Table 5:Cross\-judge SCS validation\.Entries report mean±\\pmstandard deviation on the shared valid Task\-1 subset \(n=78n=78\)\.Tab\.[5](https://arxiv.org/html/2605.15677#S5.T5)shows noticeable judge calibration differences in absolute SCS values, especially for weaker outputs\. However, all three judges produce the same model ranking: Gemini\-3\-Pro\>\>Claude\-4\.5\-Sonnet\>\>GPT\-5\.2\>\>GLM\-4\.6V\. Pairwise Spearman correlations are also consistent \(0\.770 for Gemini–GPT, 0\.735 for Gemini–Claude, and 0\.721 for GPT–Claude; average 0\.742\)\. Thus, we interpret SCS primarily as a relative visual\-quality signal: its absolute values are judge\-calibrated, but the ranking conclusions used in our analysis are stable across evaluator families\.

SigLIP2 Visual Similarity\.To capture high\-level structural and layout alignment, we compute cosine similarity between embeddings of rendered imageIg​e​nI\_\{gen\}and ground truthIg​tI\_\{gt\}using SigLIP2\(Tschannenet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib47)\):

Svis=cos⁡\(Eimg​\(Ig​e​n\),Eimg​\(Ig​t\)\)S\_\{\\text\{vis\}\}=\\cos\(E\_\{\\text\{img\}\}\(I\_\{gen\}\),E\_\{\\text\{img\}\}\(I\_\{gt\}\)\)\(2\)whereEimgE\_\{\\text\{img\}\}denotes the image encoder\. This approach captures high\-level structural and layout semantics more robustly than pixel\-level metrics\.

##### Semantic Compliance Metrics\.

CodeXQA \(Task 1\)\.We introduceCodeXQAto evaluate whether generatedmxGraphXML topologically preserves source image information\. We generate question setsQ=\{qcnt,qid,qrel\}Q=\\\{q\_\{\\text\{cnt\}\},q\_\{\\text\{id\}\},q\_\{\\text\{rel\}\}\\\}covering three cognitive levels: \(1\)Counting: enumerating elements; \(2\)Identification: retrieving attributes; and \(3\)Relationship: reasoning about connectivity\. The metric measures accuracy in answering questions based solely on the generatedmxGraphXML code\.

XDRFR \(Task 2\)\.We proposeXDRFR \(XML\-based DRFR\), adapted from DRFR\(Qinet al\.,[2024](https://arxiv.org/html/2605.15677#bib.bib46)\), to evaluate instruction following in code editing\. Each editing instructionTTis decomposed into atomic verification questionsQT=\{q1,…,qn\}Q\_\{T\}=\\\{q\_\{1\},\\dots,q\_\{n\}\\\}\. Evaluation operates directly on the modifiedmxGraphXML structure rather than rendered images, leveraging structured attribute information for precise logic verification and avoiding visual ambiguities\. The score is the ratio of satisfied requirements:

XDRFR=1\|QT\|​∑i=1\|QT\|𝕀​\(Verify​\(Ct​g​t,qi\)\)\\text\{XDRFR\}=\\frac\{1\}\{\|Q\_\{T\}\|\}\\sum\_\{i=1\}^\{\|Q\_\{T\}\|\}\\mathbb\{I\}\(\\text\{Verify\}\(C\_\{tgt\},q\_\{i\}\)\)\(3\)We conducted a manual spot\-check of 55 samples for the XDRFR metric \(App\.[A](https://arxiv.org/html/2605.15677#A1), Tab\.[11](https://arxiv.org/html/2605.15677#A1.T11)\), confirming high reliability with only one minor issue requiring manual correction in a single sample\.

## 6Experiments

### 6\.1Experimental Setup

We evaluate leading multimodal models on VCG\-Bench for baselines\. For all models we use temperature 0 and a maximum token limit \(e\.g\., 4096\) for deterministic outputs\. For Task 1, Gemini\-3\-Pro generates structured descriptions per image to assistmxGraphXML generation; for Task 2 \(Editability\), we providediffpatch schema in the prompt\. We evaluate closed\-source models \(Gemini\-3\-Pro\(Google DeepMind,[2025](https://arxiv.org/html/2605.15677#bib.bib48)\), Gemini\-3\-Flash\(Google DeepMind,[2025](https://arxiv.org/html/2605.15677#bib.bib48)\), Claude\-4\.5\-Opus\(Anthropic,[2025](https://arxiv.org/html/2605.15677#bib.bib49)\), Claude\-4\.5\-Sonnet\(Anthropic,[2025](https://arxiv.org/html/2605.15677#bib.bib49)\), GPT\-5\.2\(OpenAI,[2025](https://arxiv.org/html/2605.15677#bib.bib50)\)\) and open\-source models \(Qwen3\-VL series 235B/32B/8B\(Qwen Team,[2025a](https://arxiv.org/html/2605.15677#bib.bib51)\), Qwen3\-Omni\-30B\(Qwen Team,[2025a](https://arxiv.org/html/2605.15677#bib.bib51)\), Qwen3\-Coder\(Qwen Team,[2025b](https://arxiv.org/html/2605.15677#bib.bib63)\), GLM\-4\.6V\(GLM\-V Team,[2025](https://arxiv.org/html/2605.15677#bib.bib52)\)\(Task 1\), GLM\-4\.6 \(Task 2\), DeepSeek\-V3\.2\(DeepSeek\-AI,[2025](https://arxiv.org/html/2605.15677#bib.bib53)\), Kimi\-K2\(Kimi Team,[2025](https://arxiv.org/html/2605.15677#bib.bib54)\), MiniMax\-M2\(MiniMax\-AI,[2025](https://arxiv.org/html/2605.15677#bib.bib55)\), Step3\(Huang and StepFun Team,[2026](https://arxiv.org/html/2605.15677#bib.bib56)\), among others\)\. Additional results and breakdowns are in App\.[A](https://arxiv.org/html/2605.15677#A1)\.

### 6\.2Vision\-to\-XML Generation

Task 1 evaluates the capability of models to translate raster images into editablemxGraphXML code, serving as the gateway to the editing workflow and requiring strong perception and structural reasoning\. We report overall results first, then break down by difficulty tier and by diagrammatic reasoning primitive\.Open\-source models fail primarily on executability \(ESR\), though some retain moderate perceptual \(CodeXQA\) scores\.Tab\.[7](https://arxiv.org/html/2605.15677#S6.T7)gives overall results; Tab\.[8](https://arxiv.org/html/2605.15677#S6.T8)gives the breakdown by difficulty\. App\.[A\.1\.1](https://arxiv.org/html/2605.15677#A1.SS1.SSS1)further compares Task 1 with a representative pipeline baseline, showing that coarse layout recovery does not recover editable Draw\.io structure\.

Input ablation\.We ablate the structured caption on 100 Task\-1 examples to test whether it provides grounding beyond the image\. Removing it lowers ESR from 1\.000 to 0\.980, SCS from 0\.805 to 0\.690, and CodeXQA from 0\.930 to 0\.786 \(Tab\.[6](https://arxiv.org/html/2605.15677#S6.T6)\), confirming that captions help localize elements, layout, and relations formxGraphconversion\. We therefore use image\-plus\-caption as the main setting and image\-only generation as a harder ablation\.

Table 6:Caption ablation for Task 1\.Results are measured on a random subset of 100 examples\. The main setting uses image plus structured caption, while the ablation removes the caption and uses only the raw image\.Table 7:Overall benchmark results for Task 1: Vision\-to\-XML Generation\.Open models fail mainly on ESR, despite moderate CodeXQA for some\.Best scores in bold\.Tab\.[7](https://arxiv.org/html/2605.15677#S6.T7)shows a large gap between closed\-source and open\-source models\. Open\-source models often produce non\-executable code \(e\.g\., Qwen3\-VL\-32B/8B have ESR 0\.0000\), while closed\-source models keep ESR above 0\.84, indicating a lack of explicit Draw\.io domain knowledge on the open side\. High perceptual scores do not guarantee valid graph generation:GLM\-4\.6Vreaches CodeXQA 0\.9032 but only 0\.7471 ESR and lowSigLIP2, i\.e\., it detects elements but fails to synthesize valid structural code; this gap aligns with counting failures \(Tab\.[9](https://arxiv.org/html/2605.15677#S6.T9)\)\. Gemini\-3\-Pro is state\-of\-the\-art on style consistency \(SCS0\.7790\); most open\-source models fail on perception \(Qwen3\-VL\-32B/8B and Qwen3\-VL\-235B score 0\.0000 and 0\.1519\), withGLM\-4\.6Vleading open\-source at 0\.3433\.

Table 8:Task 1 \(Vision\-to\-XML Generation\) performance breakdown across difficulty tiers\.D\.C\.ModelSCSCodeXQAESRSigLIP2EasyClosedClaude\-4\.5\-Opus0\.71290\.94300\.94340\.8983GPT\-5\.20\.61690\.95130\.85320\.8094Gemini\-3\-Pro0\.82790\.93700\.98950\.9438Gemini\-3\-Flash0\.80940\.94360\.97890\.9316Claude\-4\.5\-Sonnet0\.69310\.94730\.96230\.9071OpenQwen3\-VL\-235B0\.22680\.92700\.45280\.3681Step30\.01080\.91700\.04610\.0334Qwen3\-VL\-8B0\.00000\.16410\.00000\.0000Qwen3\-VL\-32B0\.00000\.47040\.00000\.0000GLM\-4\.6V0\.43420\.93300\.79250\.7203Qwen3\-Omni\-30B0\.22890\.73740\.62470\.4846MediumClosedClaude\-4\.5\-Opus0\.64250\.91690\.91390\.8705GPT\-5\.20\.55510\.93710\.84580\.8037Gemini\-3\-Pro0\.77530\.94010\.96940\.9218Gemini\-3\-Flash0\.76080\.93680\.95830\.9142Claude\-4\.5\-Sonnet0\.65080\.92530\.96530\.8946OpenQwen3\-VL\-235B0\.12890\.82470\.30420\.2500Step30\.01300\.85200\.05280\.0387Qwen3\-VL\-8B0\.00000\.09360\.00000\.0000Qwen3\-VL\-32B0\.00000\.33090\.00000\.0000GLM\-4\.6V0\.32360\.90570\.74440\.6583Qwen3\-Omni\-30B0\.17710\.66840\.60830\.4659HardClosedClaude\-4\.5\-Opus0\.55270\.71310\.92000\.8545GPT\-5\.20\.47520\.84030\.81200\.7563Gemini\-3\-Pro0\.69580\.89480\.89110\.8469Gemini\-3\-Flash0\.66420\.86170\.90400\.8573Claude\-4\.5\-Sonnet0\.53470\.76840\.94400\.8398OpenQwen3\-VL\-235B0\.07550\.58570\.26400\.2032Step30\.01220\.64910\.06000\.0411Qwen3\-VL\-8B0\.00000\.04350\.00000\.0000Qwen3\-VL\-32B0\.00000\.20300\.00000\.0000GLM\-4\.6V0\.22690\.83260\.66800\.5279Qwen3\-Omni\-30B0\.17240\.41350\.70000\.5076

Across Easy, Medium, and Hard levels \(bymxGraphXML token count, Fig\.[5](https://arxiv.org/html/2605.15677#S4.F5)\), closed\-source models like Gemini\-3\-Pro show controlled degradation \(CodeXQA 0\.9370→\\to0\.8948,−4\.5%\-4\.5\\%\), whereas Qwen3\-VL\-8B/32B collapse on Hard and Qwen3\-VL\-235B decays sharply \(−36\.8%\-36\.8\\%\)\. OnlyGLM\-4\.6Vstays relatively stable, though still below closed\-source models\.

Table 9:Task 1 accuracy across diagrammatic reasoning primitives\. Evaluates models’ ability to parse structural elements from raster images\.#### 6\.2\.1Error Analysis

We analyze Task 1 failures through the diagrammatic primitives in Tab\.[9](https://arxiv.org/html/2605.15677#S6.T9)—Counting,Identification, andRelationship—which correspond to enumerating elements, recognizing types/attributes, and inferring connectivity in vision\-to\-XML\. This clarifies why open\-source models lag on executability \(Tab\.[7](https://arxiv.org/html/2605.15677#S6.T7)\) and difficulty tiers \(Tab\.[8](https://arxiv.org/html/2605.15677#S6.T8)\)\.

Counting as the dominant failure mode\.Counting is the weakness for open\-source models\. Qwen3\-VL\-8B reaches 0\.1492 \(vs\. Gemini\-3\-Pro 0\.9488\); Qwen3\-VL\-32B 0\.3064; Qwen3\-VL\-235B stays at 0\.8092\. Thisinstance\-groundingfailure drives low SCS: without accurate counts, models cannot produce topologically correct graphs, explaining the gap between moderate CodeXQA and low ESR/SCS \(e\.g\.,GLM\-4\.6Vin Tab\.[7](https://arxiv.org/html/2605.15677#S6.T7)\)\. Identification and Relationship show deficits but less severe for stronger—e\.g\.,GLM\-4\.6Vattains 0\.9326 Identification and 0\.8959 Relationship, near closed\-source; Counting \(0\.8812\) is the bottleneck\.

Open vs\. closed and within\-model patterns\.Closed\-source models stay balanced across all three primitives \(≈\\approx0\.89–0\.95\)\. Among open\-source ones, Qwen3\-VL\-8B fails uniformly \(0\.08–0\.15\); Qwen3\-VL\-32B and Qwen3\-VL\-235B show Counting as the bottleneck\. Step3 reaches 0\.8678 Identification but only 0\.7989 Counting, again highlighting instance grounding as the primary failure mode\. Fig\.[6](https://arxiv.org/html/2605.15677#S5.F6)shows qualitative examples with SCS and expert rankings\.

### 6\.3Instruction\-based Diagram Editing

Task 2 evaluates the hypothesis that, once a diagram is available as structured code, constrained incremental editing becomes substantially more precise and controllable\. We evaluate the same closed\- and open\-source models on instruction\-based patch editing; all achieve high ESR and SCS, with XDRFR distinguishing instruction\-following quality\. Tab\.[10](https://arxiv.org/html/2605.15677#S6.T10)gives the overall results\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/t2.png)Figure 7:Task 2 instruction editing precision demonstration\.Rows show distinct editing instructions; columns show the input diagram, instruction, and outputs from five representative models\.Table 10:Overall benchmark results for Task 2: Instruction\-based Diagram Editing\.High performance across models suggests that constrained code\-mediated editing is a robust paradigm\.Best scores in bold\.Under constrained XML patching, all models achieve high editing fidelity; Gemini\-3\-Pro leads with XDRFR0\.9405\. With ESR and SCS nearly saturated,XDRFRbecomes the key discriminative signal \(0\.8873–0\.9405\)\. This indicates that frontier models are mature at localized edits given complete diagram structure, but not necessarily at open\-ended redesign or ambiguous intent\. Qwen3\-VL\-235B reaches 0\.9312, ahead of some closed\-source models, while Qwen3\-VL\-32B pairs high SCS \(0\.9295\) with lower XDRFR \(0\.8873\), suggesting stronger semantic preservation than detail precision\. Fig\.[7](https://arxiv.org/html/2605.15677#S6.F7)gives qualitative examples; App\.[A\.1\.6](https://arxiv.org/html/2605.15677#A1.SS1.SSS6)compares against a pixel\-based editing baseline\.

Fig\.[7](https://arxiv.org/html/2605.15677#S6.F7)covers representative edits: text/connector updates, color changes, shape/style transformation, and multi\-operation edits over shape, position, line style, and connectors\. Across model families, outputs preserve layout and style while applying local changes, explaining high SCS and why XDRFR remains the main discriminative metric\. These cases also show XML patching’s practical advantage: edits localize to explicit nodes, edges, and attributes rather than regenerated pixels, making results easier to verify and refine in Draw\.io\.

### 6\.4Discussion

Comparing Task 1 and Task 2 \(Tabs\.[7](https://arxiv.org/html/2605.15677#S6.T7)and[10](https://arxiv.org/html/2605.15677#S6.T10)\) reveals a sharp vision\-to\-code vs\. code\-to\-code gap\. Models such asQwen3\-VL\-32Bfail on Task 1 \(generatingmxGraphXML from images\) yet achieve strong Task\-2 editing fidelity\. They can reason about XML structure and follow edits; the bottleneck isvisual understanding—inferring diagram structure from pixels\. VCG\-Bench separates these capabilities and shows that models unable to produce code from images can still edit provided code effectively\.

Scale vs\. performance\.The Qwen3\-VL series \(8B, 32B, 235B\) shows diminishing returns: perceptual gains \(CodeXQA: 0\.1041→\\to0\.8224\) outpace structured generation \(ESR: 0\.0000→\\to0\.3462, SCS: 0\.0000→\\to0\.1519\)\.GLM\-4\.6Voutperforms larger models despite fewer parameters—architecture and training data quality matter more than scalefor structured generation\.

## 7Conclusion

We have introducedVCG\-Bench, the first unified benchmark designed for evaluating controllable structured generation in VLMs\. By shifting the focus from passive perception to active, editable code generation, VCG\-Bench exposes significant limitations in current state\-of\-the\-art models in fine\-grained visual parsing and syntactic constraint satisfaction\. Our framework offers researchers a reproducible and quantifiable testbed to assess the practical readiness of multimodal agents for professional engineering workflows\. We will release the benchmark and evaluation toolkit to support future work\. Future work will expand the dataset to include dynamic interaction histories and multi\-page diagrams\.

## Acknowledgement

This paper was supported by the NSF of China \(62402409\); Youth S&T Talent Support Programme of Guangdong Provincial Association for Science and Technology \(SKXRC2025461\); the Young Talent Support Project of Guangzhou Association for Science and Technology \(QT\-2025\-001\); Guangzhou Basic and Applied Basic Research Foundation \(2026A1515010269, 2025A04J3935, 2023A1515110545\); and Guangzhou\-HKUST\(GZ\) Joint Funding Program \(2025A03J3714\); Hong Kong CRF grants under Grant No\. C7004\-22G and C6015\-23G\.

## Impact Statement

This paper presents work whose goal is to advance the field of Machine Learning, specifically in the evaluation and development of Vision\-Language Models \(VLMs\) for structured visual content generation\. VCG\-Bench introduces a standardized benchmark for assessing diagram generation and editing capabilities, which has broader implications for software engineering, technical documentation, and design automation workflows\.

We believe the primary societal impact will be positive, as improved diagram generation capabilities can enhance productivity in technical fields, improve accessibility of visual information, and support educational applications\. The open\-source nature of our benchmark promotes transparency and reproducibility in research, while the standardized evaluation protocol helps identify and address model limitations before deployment\.

## Limitations

VCG\-Bench is limited toDraw\.io/mxGraphand has not been validated on other diagram languages \(e\.g\., Visio, TikZ, Mermaid\) or in real\-world deployment\. The dataset may underrepresent non\-English labels, rare styles, and specialized notations\. The current editing task focuses on objectively verifiable, rule\-based incremental edits; it does not fully capture semantic ambiguity, mixed\-initiative clarification, or subjective diagram\-level redesign\.Executabilitydoes not guarantee correctness—diagrams can render yet contain semantic or topological errors, requiring expert review for high\-stakes\. Future work may extend to other formats, interactive clarification protocols, and real\-world suites\.

## References

- Anthropic \(2025\)Claude 4\.5 system card\.Technical reportAnthropic\.Note:Evaluation of Claude 4\.5 Sonnet and Opus modelsExternal Links:[Link](https://assets.anthropic.com/m/12f214efcc2f457a/original/Claude-Sonnet-4-5-System-Card.pdf)Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p4.1),[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- S\. Antol, A\. Agrawal, J\. Lu, M\. Mitchell, D\. Batra, C\. L\. Zitnick, and D\. Parikh \(2015\)Vqa: visual question answering\.InProceedings of the IEEE international conference on computer vision,pp\. 2425–2433\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- L\. Bai, Z\. Cai, M\. Cao, W\. Cao, C\. Chen, H\. Chen, K\. Chen, P\. Chen,et al\.\(2025\)Intern\-s1: a scientific multimodal foundation model\.arXiv preprint arXiv:2508\.15763\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- T\. Beltramelli \(2018\)Pix2code: generating code from a graphical user interface screenshot\.InProceedings of the ACM SIGCHI symposium on engineering interactive computing systems,pp\. 1–6\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- T\. Brooks, A\. Holynski, and A\. A\. Efros \(2023\)Instructpix2pix: learning to follow image editing instructions\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 18392–18402\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p5.1)\.
- A\. Carlier, M\. Danelljan, A\. Alahi, and R\. Timofte \(2020\)DeepSVG: a hierarchical generative network for vector graphics animation\.Advances in Neural Information Processing Systems33,pp\. 16351–16361\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- L\. Chen, J\. Li, X\. Dong, P\. Zhang, C\. He, J\. Wang, F\. Zhao, and D\. Lin \(2024\)Are we on the right way for evaluating large vision\-language models?\.InAdvances in Neural Information Processing Systems,Vol\.37\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- M\. Chen, J\. Tworek, H\. Jun, Q\. Yuan, H\. P\. d\. O\. Pinto, J\. Kaplan, H\. Edwards, Y\. Burda, N\. Joseph, G\. Brockman,et al\.\(2021\)Evaluating large language models trained on code\.arXiv preprint arXiv:2107\.03374\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- S\. Chen, X\. Dong, H\. Xu, X\. Wu, F\. Tang, H\. Zhang, Y\. Yan, L\. Wu, W\. Zhang, G\. Hou,et al\.\(2025a\)Svgenius: benchmarking llms in svg understanding, editing and generation\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 13289–13296\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- Z\. Chen, W\. Wang, Y\. Cao, Y\. Liu, Z\. Gao, E\. Cui, J\. Zhu, S\. Ye,et al\.\(2025b\)Expanding performance boundaries of open\-source multimodal models with model, data, and test\-time scaling\.InThe Thirteenth International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- Z\. Cui, J\. Yuan, H\. Wang, Y\. Li, C\. Du, and Z\. Ding \(2025\)Draw with thought: unleashing multimodal reasoning for scientific diagram generation\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 5050–5059\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- DeepSeek\-AI \(2025\)DeepSeek\-v3\.2: pushing the frontier of open large language models\.arXiv preprint arXiv:2512\.02556\.External Links:[Link](https://arxiv.org/abs/2512.02556)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- P\. Dong, Z\. Tang, X\. Liu, L\. Li, X\. Chu, and B\. Li \(2025\)Can compressed LLMs truly act? An empirical evaluation of agentic capabilities in LLM compression\.InProceedings of the 42nd International Conference on Machine Learning,Cited by:[Appendix F](https://arxiv.org/html/2605.15677#A6.p1.1)\.
- GLM\-V Team \(2025\)GLM\-4\.1V\-Thinking: towards versatile multimodal reasoning with scalable reinforcement learning\.arXiv preprint arXiv:2507\.01006\.External Links:[Link](https://arxiv.org/abs/2507.01006)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- Google DeepMind \(2025\)A new era of intelligence with gemini 3\.Technical reportGoogle DeepMind\.Note:Gemini 3 Technical Report and Model ReleaseExternal Links:[Link](https://blog.google/products-and-platforms/products/gemini/gemini-3/)Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1),[§1](https://arxiv.org/html/2605.15677#S1.p4.1),[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- D\. Guo, Q\. Zhu, D\. Yang, Z\. Xie, K\. Dong, W\. Zhang, G\. Chen, X\. Bi, Y\. Wu, Y\. Li,et al\.\(2024\)DeepSeek\-coder: when the large language model meets programming–the rise of code intelligence\.arXiv preprint arXiv:2401\.14196\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- M\. He, J\. Zeng, Y\. Jiang, W\. Zhang, Z\. Liu, X\. Shi, and A\. Zhou \(2025\)Flow2Code: evaluating large language models for flowchart\-based code generation capability\.arXiv preprint arXiv:2506\.02073\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- A\. Hertz, R\. Mokady, J\. Tenenbaum, K\. Aberman, Y\. Pritch, and D\. Cohen\-Or \(2023\)Prompt\-to\-prompt image editing with cross attention control\.InThe Eleventh International Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p5.1)\.
- A\. Huang and StepFun Team \(2026\)STEP3\-vl\-10b technical report\.arXiv preprint arXiv:2601\.09668\.External Links:[Link](https://arxiv.org/abs/2601.09668)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- A\. Jain, A\. Xie, and P\. Abbeel \(2023\)Vectorfusion: text\-to\-svg by abstracting pixel\-based diffusion models\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 1911–1920\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- V\. Jain, P\. Agrawal, S\. Banga, R\. Kapoor, and S\. Gulyani \(2019\)Sketch2Code: transformation of sketches to ui in real\-time using deep neural network\.arXiv preprint arXiv:1910\.08930\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- S\. E\. Kahou, V\. Michalski, A\. Atkinson, Á\. Kádár, A\. Trischler, and Y\. Bengio \(2018\)FigureQA: an annotated figure dataset for visual reasoning\.InInternational Conference on Learning Representations Workshop,Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- A\. Kembhavi, M\. Salvato, E\. Kolve, M\. Seo, H\. Hajishirzi, and A\. Farhadi \(2016\)A diagram is worth a dozen images\.InEuropean conference on computer vision,pp\. 235–251\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- Kimi Team \(2025\)Kimi k2: open agentic intelligence\.arXiv preprint arXiv:2507\.20534\.External Links:[Link](https://arxiv.org/abs/2507.20534)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- B\. Li, Y\. Ge, Y\. Ge, G\. Wang, R\. Wang, R\. Zhang, and Y\. Shan \(2024a\)Seed\-bench: benchmarking multimodal large language models\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 13299–13308\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- K\. Li, Y\. Tian, Q\. Hu, Z\. Luo, Z\. Huang, and J\. Ma \(2024b\)Mmcode: benchmarking multimodal large language models for code generation with visually rich programming problems\.arXiv preprint arXiv:2404\.09486\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.6.2.2.3),[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- R\. Li, L\. B\. Allal, Y\. Zi, N\. Muennighoff, D\. Kocetkov, C\. Mou, M\. Marone, C\. Akiki, J\. Li, J\. Chim,et al\.\(2023\)Starcoder: may the source be with you\!\.arXiv preprint arXiv:2305\.06161\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- Z\. Li, X\. Yang, K\. Choi, W\. Zhu, R\. Hsieh, H\. Kim, J\. H\. Lim, S\. Ji, B\. Lee, X\. Yan,et al\.\(2024c\)Mmsci: a multimodal multi\-discipline dataset for phd\-level scientific comprehension\.InAI for Accelerated Materials Design\-Vienna 2024,Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- Z\. Liao, K\. Jiang, Z\. Liu, Y\. Wei, J\. Yu, Q\. Li, H\. Yu, P\. Li,et al\.\(2026\)AIBench: evaluating visual\-logical consistency in academic illustration generation\.arXiv preprint arXiv:2603\.28068\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- K\. Q\. Lin, Y\. Zheng, H\. Ran, D\. Zhu, D\. Mao, L\. Li, P\. Torr, and A\. J\. Wang \(2025\)VCode: a multimodal coding benchmark with svg as symbolic visual representation\.arXiv preprint arXiv:2511\.02778\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.14.10.10.3),[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- H\. Liu, C\. Li, Y\. Li, and Y\. J\. Lee \(2023\)Improved baselines with visual instruction tuning\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- X\. Liu, P\. Dong, X\. Hu, and X\. Chu \(2024a\)LongGenBench: long\-context generation benchmark\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Cited by:[Appendix F](https://arxiv.org/html/2605.15677#A6.p1.1)\.
- X\. Liu, Z\. Tang, H\. Chen, P\. Dong, Z\. Li, X\. Zhou, B\. Li, X\. Hu, and X\. Chu \(2026\)Can LLMs maintain fundamental abilities under KV cache compression?\.InProceedings of the 43rd International Conference on Machine Learning,Cited by:[Appendix F](https://arxiv.org/html/2605.15677#A6.p1.1)\.
- X\. Liu, Z\. Tang, P\. Dong, Z\. Li, B\. Li, X\. Hu, and X\. Chu \(2025\)ChunkKV: semantic\-preserving KV cache compression for efficient long\-context LLM inference\.InAdvances in Neural Information Processing Systems,Cited by:[§C\.3](https://arxiv.org/html/2605.15677#A3.SS3.p2.1)\.
- Y\. Liu, H\. Duan, Y\. Zhang, B\. Li, S\. Zhang, W\. Zhao, Y\. Yuan, J\. Wang, C\. He, Z\. Liu,et al\.\(2024b\)Mmbench: is your multi\-modal model an all\-around player?\.InEuropean conference on computer vision,pp\. 216–233\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.22.18.18.3),[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- A\. Lozhkov, R\. Li, L\. Ben Allal, F\. Cassano, J\. Lamy\-Poirier, N\. Tazi, A\. Tang, D\. Pykhtar,et al\.\(2024\)StarCoder 2 and the stack v2: the next generation\.arXiv preprint arXiv:2402\.19173\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- A\. Masry, D\. X\. Long, J\. Q\. Tan, S\. Joty, and E\. Hoque \(2022\)ChartQA: a benchmark for question answering about charts with visual and logical reasoning\.InFindings of the Association for Computational Linguistics: ACL 2022,S\. Muresan, P\. Nakov, and A\. Villavicencio \(Eds\.\),Dublin, Ireland,pp\. 2263–2279\.External Links:[Link](https://aclanthology.org/2022.findings-acl.177/),[Document](https://dx.doi.org/10.18653/v1/2022.findings-acl.177)Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- MiniMax\-AI \(2025\)MiniMax\-m2: efficient agentic intelligence\.Hugging Face\.Note:Open Source Release for Agentic Coding WorkflowsExternal Links:[Link](https://huggingface.co/MiniMaxAI/MiniMax-M2)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- K\. Nishina and Y\. Matsui \(2024\)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 Workshops,pp\. 8142–8147\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- K\. Nishina and Y\. Matsui \(2025\)Svgeditbench v2: a benchmark for instruction\-based svg editing\.arXiv preprint arXiv:2502\.19453\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- T\. Niu, Y\. Cui, B\. Wang, X\. Xu, X\. Yao, Q\. Zhu, D\. Wu, S\. Wang, and W\. Che \(2025\)Chart2Code53: a large\-scale diverse and complex dataset for enhancing chart\-to\-code generation\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 15839–15855\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- OpenAI \(2023\)GPT\-4 technical report\.arXiv preprint arXiv:2303\.08774\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- OpenAI \(2025\)Introducing gpt\-5\.2: specialist\-level reasoning and reliability\.Note:Technical overview of GPT\-5\.2 Thinking and Pro modelsExternal Links:[Link](https://openai.com/index/introducing-gpt-5-2/)Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p4.1),[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- OpenDCAI \(2025\)Paper2Any\.External Links:[Link](https://github.com/OpenDCAI/Paper2Any)Cited by:[§A\.1\.1](https://arxiv.org/html/2605.15677#A1.SS1.SSS1.p1.1)\.
- H\. Pan, Q\. Zhang, C\. Caragea, E\. Dragut, and L\. J\. Latecki \(2024\)Flowlearn: evaluating large vision\-language models on flowchart understanding\.arXiv preprint arXiv:2407\.05183\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- D\. Podell, Z\. English, K\. Lacey, A\. Blattmann, T\. Dockhorn, J\. Müller, J\. Penna, and R\. Rombach \(2024\)SDXL: improving latent diffusion models for high\-resolution image synthesis\.InThe Twelfth International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- Y\. Qin, K\. Song, Y\. Hu, W\. Yao, S\. Cho, X\. Wang, X\. Wu, F\. Liu, P\. Liu, and D\. Yu \(2024\)Infobench: evaluating instruction following ability in large language models\.arXiv preprint arXiv:2401\.03601\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1),[§5](https://arxiv.org/html/2605.15677#S5.SS0.SSS0.Px3.p2.2)\.
- Qwen Team \(2025a\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.External Links:[Link](https://arxiv.org/abs/2505.09388)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- Qwen Team \(2025b\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.External Links:[Link](https://arxiv.org/abs/2505.09388)Cited by:[§6\.1](https://arxiv.org/html/2605.15677#S6.SS1.p1.1)\.
- P\. Reddy, M\. Gharbi, M\. Lukac, and N\. J\. Mitra \(2021\)Im2Vec: synthesizing vector graphics without vector supervision\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 7342–7351\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- R\. Rombach, A\. Blattmann, D\. Lorenz, P\. Esser, and B\. Ommer \(2022\)High\-resolution image synthesis with latent diffusion models\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 10684–10695\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1),[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- B\. Roziere, J\. Gehring, F\. Gloeckle, S\. Sootla, I\. Gat, X\. E\. Tan, Y\. Adi, J\. Liu, R\. Sauvestre, T\. Remez,et al\.\(2023\)Code llama: open foundation models for code\.arXiv preprint arXiv:2308\.12950\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- C\. Si, Y\. Zhang, R\. Li, Z\. Yang, R\. Liu, and D\. Yang \(2025\)Design2code: benchmarking multimodal code generation for automated front\-end engineering\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 3956–3974\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.12.8.8.3),[§2](https://arxiv.org/html/2605.15677#S2.p4.1),[§2](https://arxiv.org/html/2605.15677#S2.p6.1),[§3\.1](https://arxiv.org/html/2605.15677#S3.SS1.p1.1)\.
- M\. Tschannen, A\. Gritsenko, X\. Wang, M\. F\. Naeem, I\. Alabdulmohsin, N\. Parthasarathy, T\. Evans, L\. Beyer, Y\. Xia, B\. Mustafa, O\. Hénaff, J\. Harmsen, A\. Steiner, and X\. Zhai \(2025\)SigLIP 2: multilingual vision\-language encoders with improved semantic understanding, localization, and dense features\.External Links:2502\.14786,[Link](https://arxiv.org/abs/2502.14786)Cited by:[§3\.2](https://arxiv.org/html/2605.15677#S3.SS2.p1.4),[§5](https://arxiv.org/html/2605.15677#S5.SS0.SSS0.Px2.p4.2)\.
- Y\. Vinker, E\. Pajouheshgar, J\. Y\. Bo, R\. C\. Bachmann, A\. H\. Bermano, D\. Cohen\-Or, A\. Zamir, and A\. Shamir \(2022\)CLIPasso: semantically\-aware object sketching\.ACM Transactions on Graphics \(TOG\)41\(4\),pp\. 1–11\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- P\. Wang, S\. Bai, S\. Tan, S\. Wang, Z\. Fan, J\. Bai, K\. Chen, X\. Liu,et al\.\(2024a\)Qwen2\-VL: enhancing vision\-language model’s perception of the world at any resolution\.InAdvances in Neural Information Processing Systems,Vol\.37\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- Z\. Wang, M\. Xia, L\. He, H\. Chen, Y\. Liu, R\. Zhu, K\. Liang, X\. Wu, H\. Liu, S\. Malladi,et al\.\(2024b\)Charxiv: charting gaps in realistic chart understanding in multimodal llms\.Advances in Neural Information Processing Systems37,pp\. 113569–113697\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.18.14.14.3),[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- J\. Wei, C\. Tan, Q\. Chen, G\. Wu, S\. Li, Z\. Gao, L\. Sun, B\. Yu, and R\. Guo \(2025\)From words to structured visuals: a benchmark and framework for text\-to\-diagram generation and editing\.InProceedings of the Computer Vision and Pattern Recognition Conference,pp\. 13315–13325\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- C\. Wu, J\. Li, J\. Zhou, J\. Lin, K\. Gao, K\. Yan, S\. Yin, S\. Bai, X\. Xu, Y\. Chen,et al\.\(2025a\)Qwen\-image technical report\.arXiv preprint arXiv:2508\.02324\.Cited by:[§1](https://arxiv.org/html/2605.15677#S1.p1.1)\.
- C\. Wu, Z\. Liang, Y\. Ge, Q\. Guo, Z\. Lu, J\. Wang, Y\. Shan, and P\. Luo \(2025b\)Plot2code: a comprehensive benchmark for evaluating multi\-modal large language models in code generation from scientific plots\.InFindings of the Association for Computational Linguistics: NAACL 2025,pp\. 3006–3028\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.10.6.6.3),[§2](https://arxiv.org/html/2605.15677#S2.p2.1),[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- R\. Wu, W\. Su, K\. Ma, and J\. Liao \(2023\)IconShop: text\-guided vector icon synthesis with autoregressive transformers\.ACM Transactions on Graphics \(TOG\)42\(6\),pp\. 1–14\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- Y\. Xie, Y\. Luo, G\. Li, and N\. Tang \(2024\)HAIChart: human and ai paired visualization system\.Proceedings of the VLDB Endowment17\(11\),pp\. 3178–3191\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- Y\. Xie, Z\. Zhang, Y\. Wu, S\. Lu, J\. Zhang, Z\. Yu, J\. Wang, S\. Hong, B\. Liu, C\. Wu,et al\.\(2026\)VisJudge\-Bench: aesthetics and quality assessment of visualizations\.InThe Fourteenth International Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- P\. Xu, W\. Shao, K\. Zhang, P\. Gao, S\. Liu, M\. Lei, F\. Meng, S\. Huang, Y\. Qiao, and P\. Luo \(2024\)Lvlm\-ehub: a comprehensive evaluation benchmark for large vision\-language models\.IEEE Transactions on Pattern Analysis and Machine Intelligence\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- C\. Yang, C\. Shi, Y\. Liu, B\. Shui, J\. Wang, M\. Jing, L\. Xu, X\. Zhu, S\. Li, Y\. Zhang, G\. Liu, X\. Nie, D\. Cai, and Y\. Yang \(2025\)ChartMimic: evaluating LMM’s cross\-modal reasoning capability via chart\-to\-code generation\.InThe Thirteenth International Conference on Learning Representations,Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.8.4.4.3),[§2](https://arxiv.org/html/2605.15677#S2.p2.1),[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- W\. Yu, Z\. Yang, L\. Li, J\. Wang, K\. Lin, Z\. Liu, X\. Wang, and L\. Wang \(2023\)MM\-Vet: evaluating large multimodal models for integrated capabilities\.arXiv preprint arXiv:2308\.02490\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.20.16.16.3),[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- X\. Yue, Y\. Ni, K\. Zhang, T\. Zheng, R\. Liu, G\. Zhang, S\. Stevens, D\. Jiang, W\. Ren, Y\. Sun,et al\.\(2024\)Mmmu: a massive multi\-discipline multimodal understanding and reasoning benchmark for expert agi\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 9556–9567\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- X\. Yue, T\. Zheng, Y\. Ni, Y\. Wang, K\. Zhang, S\. Tong, Y\. Sun, B\. Yu, G\. Zhang, H\. Sun,et al\.\(2025\)Mmmu\-pro: a more robust multi\-discipline multimodal understanding benchmark\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 15134–15186\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- J\. Zhang, Y\. Li, Z\. Li, X\. Guo, J\. Wu, L\. Zheng, Y\. Yang, J\. Zhang,et al\.\(2026\)RealChart2Code: advancing chart\-to\-code generation with real data and multi\-task evaluation\.arXiv preprint arXiv:2603\.25804\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p2.1)\.
- J\. Zhang, J\. Zhang, Z\. Cui, J\. Yang, L\. Zhang, B\. Hui, Q\. Liu, Z\. Wang, L\. Wang, and J\. Lin \(2025\)PlotCraft: pushing the limits of llms for complex and interactive data visualization\.arXiv preprint arXiv:2511\.00010\.Cited by:[Table 2](https://arxiv.org/html/2605.15677#S2.T2.16.12.12.3),[§2](https://arxiv.org/html/2605.15677#S2.p2.1),[§2](https://arxiv.org/html/2605.15677#S2.p3.1)\.
- K\. Zhang, L\. Mo, W\. Chen, H\. Sun, and Y\. Su \(2023a\)Magicbrush: a manually annotated dataset for instruction\-guided image editing\.Advances in Neural Information Processing Systems36,pp\. 31428–31449\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p5.1)\.
- L\. Zhang, A\. Rao, and M\. Agrawala \(2023b\)Adding conditional control to text\-to\-image diffusion models\.InProceedings of the IEEE/CVF international conference on computer vision,pp\. 3836–3847\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p4.1)\.
- J\. Zhou, T\. Lu, S\. Mishra, S\. Brahma, S\. Basu, Y\. Luan, D\. Zhou, and L\. Hou \(2023\)Instruction\-following evaluation for large language models\.arXiv preprint arXiv:2311\.07911\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.
- L\. Zhuo, S\. Han, Y\. Pu, B\. Qiu, S\. Paul, Y\. Liao, Y\. Liu, J\. Shao, X\. Chen, S\. Liu,et al\.\(2025\)Factuality matters: when image generation and editing meet structured visuals\.arXiv preprint arXiv:2510\.05091\.Cited by:[§2](https://arxiv.org/html/2605.15677#S2.p6.1)\.

Appendix Contents

AExperimental Results\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[A](https://arxiv.org/html/2605.15677#A1)

A\.1Additional Experimental Results\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[A](https://arxiv.org/html/2605.15677#A1)

A\.2Paper2Any Pipeline Baseline for Task 1\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[A\.1\.1](https://arxiv.org/html/2605.15677#A1.SS1.SSS1)

BDataset Details\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B](https://arxiv.org/html/2605.15677#A2)

B\.1Dataset Taxonomy\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B\.1](https://arxiv.org/html/2605.15677#A2.SS1)

B\.2Data Source Details\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B\.2](https://arxiv.org/html/2605.15677#A2.SS2)

B\.3Data Synthesis Pipeline\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B\.3](https://arxiv.org/html/2605.15677#A2.SS3)

B\.4Dataset Statistics and Difficulty\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B\.4](https://arxiv.org/html/2605.15677#A2.SS4)

B\.5Dataset and Licensing\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[B\.5](https://arxiv.org/html/2605.15677#A2.SS5)

CAdditional Information\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[C](https://arxiv.org/html/2605.15677#A3)

C\.1Annotation Schema Details\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[C](https://arxiv.org/html/2605.15677#A3)

C\.2Ethical Considerations and Release\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[C](https://arxiv.org/html/2605.15677#A3)

C\.3Discussion\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[C](https://arxiv.org/html/2605.15677#A3)

DTechnical Specifications\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[D](https://arxiv.org/html/2605.15677#A4)

D\.1mxGraph XML Schema Overview\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[D](https://arxiv.org/html/2605.15677#A4)

D\.2Comparative Analysis of Visual Representation Formats\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[D](https://arxiv.org/html/2605.15677#A4)

EPrompt Templates and Instructions\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E](https://arxiv.org/html/2605.15677#A5)

E\.1Task 1: Diagram Generation Prompts\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E](https://arxiv.org/html/2605.15677#A5)

E\.2Task 2: Diagram Editing Prompts\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E](https://arxiv.org/html/2605.15677#A5)

E\.3Evaluation\-Related Prompts\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E](https://arxiv.org/html/2605.15677#A5)

E\.4Instruction Templates\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[E](https://arxiv.org/html/2605.15677#A5)

FEvaluation Methods and Metrics\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F](https://arxiv.org/html/2605.15677#A6)

F\.1Question Set Construction\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F](https://arxiv.org/html/2605.15677#A6)

F\.2Evaluation Metrics Details\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F](https://arxiv.org/html/2605.15677#A6)

F\.3Validation and Reproducibility\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.\.[F](https://arxiv.org/html/2605.15677#A6)

## Appendix AExperimental Results

### A\.1Additional Experimental Results

#### A\.1\.1Paper2Any Pipeline Baseline for Task 1

We additionally examine Paper2Any\(OpenDCAI,[2025](https://arxiv.org/html/2605.15677#bib.bib64)\)as a representative non\-LLM pipeline baseline for Task 1\. This baseline is closer to an OCR, detection, layout analysis, and rule\-based recovery pipeline than to an end\-to\-end VLM\. The purpose of this comparison is to contextualize whether Task 1 mainly tests XML formatting or the harder problem of recovering editable diagram structure\.

In our qualitative comparison, Paper2Any can produce visual outputs, but its decomposition is often coarse\. It may treat the whole diagram as a single block or partition it into a small number of rectangular regions, producing slide\-like patches rather than editable Draw\.io objects\. As a result, the recovered output lacks fine\-grained nodes, edges, connectors, and topology that can be directly manipulated inmxGraph\. This supports the claim that Task 1 evaluates internal structure and editable semantic recovery, rather than merely converting an input into an XML\-like file format\.

We include the input diagrams, generated outputs, and visual comparisons for this baseline in the released benchmark repository together with the other supplementary experimental artifacts\.

#### A\.1\.2Performance Trade\-offs on the Pareto Front

Figure[8](https://arxiv.org/html/2605.15677#A1.F8)visualizes the performance trade\-offs between visual consistency \(measured by SCS on Task 1\) and editing fidelity \(measured by XDRFR on Task 2\) across frontier model families \(GPT, Claude, Gemini\)\. The scatter plot reveals several key insights: \(1\)Performance correlation: The positive correlation between SCS and XDRFR suggests that the underlying capabilities for structured understanding and code manipulation translate across tasks, validating that the code\-mediated paradigm enables models to maintain both high visual fidelity and precise editability\. Models from the Gemini, GPT, and Claude families tend to achieve strong performance on both dimensions simultaneously, indicating that strong visual code generation capabilities \(Task 1\) correlate with precise editing capabilities \(Task 2\)\. \(2\)Model family characteristics: Different model families may exhibit distinct positioning patterns, with some models potentially excelling in one dimension while maintaining competitive performance in the other\. This analysis is crucial for understanding the transferability of capabilities between visual code generation and editing tasks, and demonstrates that the code\-based representation provides a unified framework for both generation and manipulation of diagrammatic content\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/x2.png)Figure 8:Performance trade\-offs\. This scatter plot compares the performance of frontier model families \(GPT, Claude, Gemini\) in terms of visual consistency \(SCS for Task 1\) and editing fidelity \(XDRFR for Task 2\)\.
#### A\.1\.3Model Robustness Across Task Difficulty Levels

Figure[9](https://arxiv.org/html/2605.15677#A1.F9)presents a heatmap displaying CodeXQA accuracy scores for 11 evaluated models across Easy, Medium, and Hard difficulty levels on Task 1\. The heatmap reveals several critical patterns: \(1\)Top\-tier performance: The top 6\-7 models \(Gemini\-3\-pro, Gemini\-3\-flash, GPT\-5\.2, GLM\-4\.6V, Claude\-sonnet\-4\-5, Claude\-opus\-4\-5\) consistently achieve high accuracy scores, generally above 0\.90 for Easy and Medium tasks, and above 0\.70 for Hard tasks\. Notably, GPT\-5\.2 achieves the highest Easy score \(0\.951\), while Gemini\-3\-pro leads on Medium \(0\.940\) and Hard \(0\.895\) tasks\. \(2\)Superior robustness of closed\-source models: These top\-performing models demonstrate strong robustness with relatively small performance degradation as task difficulty increases\. For instance, Gemini\-3\-pro’s score only drops from 0\.937 \(Easy\) to 0\.895 \(Hard\), a decline of only 4\.2 percentage points\. \(3\)Significant degradation for smaller models: Models like Qwen3\-Omni\-30B, Qwen3\-VL\-32B, and Qwen3\-VL\-8B show sharp performance declines with increasing difficulty\. Qwen3\-VL\-8B performs particularly poorly, dropping from 0\.164 \(Easy\) to 0\.044 \(Hard\), representing a 73% relative decrease\. \(4\)Widening performance gap: The difference between top\-tier and lower\-tier models becomes substantially wider as task difficulty increases, highlighting that complex diagrams require advanced reasoning capabilities currently concentrated in frontier models\. This analysis validates the importance of difficulty stratification in our benchmark and reveals the current limitations of open\-source alternatives in handling sophisticated visual code generation tasks\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/task1_academic_heatmap.png)Figure 9:Model robustness across task difficulty levels\. The heatmap shows CodeXQA accuracy scores of evaluated models on Easy, Medium, and Hard samples\. Closed\-source models demonstrate superior stability, while open\-source and smaller\-scale models show sharp performance degradation\.
#### A\.1\.4XDRFR Human Correction Audit

To ensure the accuracy and reliability of the XDRFR evaluation metric, we conducted a manual audit of the automated evaluation results\. We randomly sampled 100 evaluation samples \(approximately 10% of the total 1,005 instructions\) and manually reviewed the Yes/No answers generated by the automated evaluation model \(gemini\-3\-pro\-preview\)\.

We selected 55 representative examples for detailed review\. Each sample is uniquely identified by four dimensions: model name, domain, sample ID, and instruction\. Among these 55 samples, onlyone sample\(sample\_0087\) required manual correction of the automated evaluation answers, demonstrating the high reliability of our automated XDRFR evaluation pipeline\. The corrected answers are highlighted inredin Table[11](https://arxiv.org/html/2605.15677#A1.T11)\.

Table 11:XDRFR Human Correction Audit: 55 Representative Samplessample\_0087sample\_0008sample\_0035sample\_0056sample\_0078GLM\-4\.6claude\-opus\-4\-5gemini\-3\-flashgpt\-5\.2claude\-sonnet\-4\-5Move Denoise box…Change Start to redDelete Intermediate StepChange arrow to blue…Add Analysis node…Q: 5, Manual CorrectionQ: 2, CorrectQ: 1, CorrectQ: 3, CorrectQ: 3, Correctsample\_0092sample\_0115sample\_0134sample\_0156sample\_0023qwen3\-vl\-235bglm\-4\.6kimi\-k2minimax\-m2qwen3\-coderChange central node…Move Navigation…Remove arrow…Change Process text…Change User to yellowQ: 4, CorrectQ: 2, CorrectQ: 1, CorrectQ: 3, CorrectQ: 2, Correctsample\_0067sample\_0089sample\_0101sample\_0124sample\_0145deepseek\-v3\.2claude\-opus\-4\-5gemini\-3\-progpt\-5\.2qwen3\-vl\-32bAdd Filter node…Change Revenue border…Move Manager down…Change Button shape…Delete connection…Q: 3, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 1, Correctsample\_0167sample\_0189sample\_0203sample\_0221sample\_0238glm\-4\.6kimi\-k2claude\-sonnet\-4\-5gemini\-3\-flashqwen3\-vl\-235bChange Payment text…Change arrow to double…Change Main Topic size…Change Task 1 color…Change Neuron shape…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, Correctsample\_0195sample\_0256sample\_0267sample\_0278sample\_0289deepseek\-v3\.2gemini\-3\-proclaude\-opus\-4\-5gpt\-5\.2qwen3\-vl\-32bChange Customer bg…Change Sales text…Delete Review node…Change Login shape…Add Cache node…Q: 3, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 3, Correctsample\_0301sample\_0312sample\_0323sample\_0334sample\_0345kimi\-k2glm\-4\.6minimax\-m2qwen3\-coderclaude\-sonnet\-4\-5Change arrow to double…Change Main Topic size…Change SWOT border…Move Phase 2 down…Change Chart to yellow…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, Correctsample\_0356sample\_0367sample\_0378sample\_0389sample\_0401gemini\-3\-flashdeepseek\-v3\.2qwen3\-vl\-235bgpt\-5\.2claude\-opus\-4\-5Change Interface shape…Change Feature A text…Change Layer shape…Change Manager shape…Change Header bg…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, Correctsample\_0412sample\_0423sample\_0434sample\_0445sample\_0456gemini\-3\-prokimi\-k2glm\-4\.6qwen3\-vl\-32bminimax\-m2Add Order table…Change Q1 Report shape…Change arrow to dashed…Delete Approval step…Change Branch 1 color…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 1, CorrectQ: 2, Correctsample\_0467sample\_0478sample\_0489sample\_0501sample\_0512claude\-sonnet\-4\-5qwen3\-codergemini\-3\-flashdeepseek\-v3\.2gpt\-5\.2Change arrow to blue…Change Goal shape…Move Axis left…Change Menu shape…Change Task 5 border…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, Correctsample\_0523sample\_0534sample\_0545sample\_0556sample\_0567claude\-opus\-4\-5qwen3\-vl\-235bkimi\-k2glm\-4\.6minimax\-m2Change Activation shape…Add Utility class…Change Module B text…Change Team Lead shape…Delete connection…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 1, Correctsample\_0578sample\_0589sample\_0601sample\_0612sample\_0623claude\-sonnet\-4\-5qwen3\-vl\-32bgemini\-3\-prodeepseek\-v3\.2gpt\-5\.2Change Footer bg…Change Sub\-branch 2…Add Product table…Change Annual Report shape…Move Verification…Q: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectQ: 2, CorrectThe sample requiring correction \(sample\_0087\) is highlighted inred\. This sample had two questions where the automated evaluation model \(gemini\-3\-pro\-preview\) incorrectly answered “No” when the correct answer should have been “Yes”\. The extremely low correction rate \(1 out of 55 samples, or 1\.8%\) demonstrates the high reliability of our automated XDRFR evaluation pipeline\.

#### A\.1\.5Detailed Performance Breakdown for Task 2 Across Difficulty Levels

Table[12](https://arxiv.org/html/2605.15677#A1.T12)presents the performance breakdown for Task 2 \(instruction\-to\-patch editing\) stratified by difficulty level\. This table evaluates three key metrics: Execution Success Rate \(ESR\), Style Consistency Score \(SCS\), and XML Decomposed Requirements Following Ratio \(XDRFR\)\.

Table 12:Task 2 performance breakdown across difficulty levelsD\.C\.ModelESRSCSXDRFREasyClosedClaude\-4\.5\-Opus1\.00000\.89890\.9254GPT\-5\.21\.00000\.90710\.9428Gemini\-3\-Pro1\.00000\.90550\.9466Gemini\-3\-Flash1\.00000\.89980\.9344Claude\-4\.5\-Sonnet1\.00000\.88320\.9269OpenKimi\-K20\.99600\.88860\.9259GLM\-4\.60\.99600\.87790\.9245Qwen3\-VL\-235B0\.99800\.87220\.9302MiniMax\-M20\.98990\.86930\.9213Qwen3\-VL\-32B1\.00000\.93190\.8841Qwen3\-Coder0\.99390\.89070\.9147DeepSeek\-V3\.21\.00000\.88990\.8911MediumClosedClaude\-4\.5\-Opus0\.99710\.92100\.9256GPT\-5\.20\.99710\.91890\.9214Gemini\-3\-Pro0\.99850\.90930\.9379Gemini\-3\-Flash0\.99850\.91540\.9308Claude\-4\.5\-Sonnet0\.99850\.89520\.9244OpenKimi\-K20\.99710\.91380\.9290GLM\-4\.60\.99120\.89950\.9182Qwen3\-VL\-235B0\.99560\.91510\.9311MiniMax\-M20\.99120\.91140\.9045Qwen3\-VL\-32B0\.99850\.92420\.8807Qwen3\-Coder0\.99270\.91270\.9293DeepSeek\-V3\.20\.99260\.91670\.8884HardClosedClaude\-4\.5\-Opus1\.00000\.92860\.9336GPT\-5\.20\.99510\.92250\.9526Gemini\-3\-Pro1\.00000\.92290\.9389Gemini\-3\-Flash1\.00000\.93680\.9188Claude\-4\.5\-Sonnet1\.00000\.91390\.9427OpenKimi\-K20\.99520\.89010\.9263GLM\-4\.61\.00000\.93890\.9194Qwen3\-VL\-235B0\.99510\.92830\.9330MiniMax\-M21\.00000\.91600\.9350Qwen3\-VL\-32B0\.99520\.94060\.9097Qwen3\-Coder1\.00000\.91860\.9076DeepSeek\-V3\.20\.99520\.92060\.9071

Key Observations:

- •Execution Success: Nearly all models achieve near\-perfect ESR \(≥\\geq0\.99\) across difficulty levels, confirming that the incremental modification format produces syntactically valid XML code reliably\.
- •Style Preservation: SCS scores remain consistently high \(0\.87\-0\.94\) across all difficulty levels, indicating that models successfully maintain visual style and aesthetic quality even when performing complex edits\. Closed\-source models \(especially Gemini\-3\-Flash and Claude\-4\.5\-Opus\) excel in style consistency\.
- •Instruction Following: XDRFR scores are uniformly high \(0\.88\-0\.95\), significantly outperforming Task 1 performance, which validates the effectiveness of our proposed incremental modification strategy and the inherent controllability of the code\-mediated paradigm\. The decomposition\-based evaluation enables precise verification of instruction compliance\.
- •Difficulty Scaling: Unlike Task 1, Task 2 performance shows remarkable stability across difficulty levels, with minimal degradation on Hard samples\. This suggests that once diagrams are represented as code, editing operations become more tractable regardless of initial complexity\.
- •Model Comparison: Closed\-source models maintain a slight edge, but open\-source alternatives \(particularly Qwen3\-VL\-235B, Kimi\-K2, and GLM\-4\.6\) demonstrate competitive performance, narrowing the gap compared to Task 1\.

Bold values indicate the best performers in each category \(closed\-source/open\-source\) for each difficulty level and metric\.

#### A\.1\.6Pixel\-based Editing Baseline for Task 2

To empirically compare the Diagram\-as\-Code editing paradigm with a Diagram\-as\-Pixels alternative, we add a pixel\-based editing baseline on a subset aligned with Task 2\. The subset contains 32 diagrams, each paired with three editing instructions at the Easy, Medium, and Hard levels, resulting in 96 editing instructions in total\. Both routes use the same English editing instructions\. For the pixel\-editing route, the input is the rendered PNG diagram and the editing instruction, and the output edited image is evaluated by human DRFR scoring over 335 decomposed questions\. For the symbolic route, we follow the XML patch evaluation protocol used in the main paper and compute XDRFR on the same sampled instructions\.

Table 13:Pixel\-based editing baseline versus symbolic XML editing on a Task\-2\-aligned subset\. DRFR is obtained from human evaluation of edited images, while XDRFR evaluates instruction compliance on the edited XML\.The results show that the pixel\-based baseline is consistently weaker than the symbolic XML editing route, both overall and across all difficulty levels\. The gap is most pronounced on Hard instructions, where the pixel baseline drops to 0\.649 DRFR, while XML\-based editing maintains 0\.867 XDRFR\. Qualitatively, pixel editing often introduces blurred or incorrectly modified text, unintentionally deletes or changes irrelevant elements, and struggles to preserve connectors and topology\. Even when the requested edit is partially applied, geometric positions, font sizes, and colors are often difficult to control precisely\. Fine\-grained local edits, such as slightly enlarging an object or moving it a little to the left, are especially unreliable in pixel space\. In contrast, the XML\-based route applies localized patches, re\-renders through the same evaluation pipeline, and preserves the option of opening the result in Draw\.io for further structured refinement\. This makes code\-mediated editing better aligned with professional diagram\-editing workflows\.

## Appendix BDataset Details

### B\.1Dataset Taxonomy

The VCG\-Bench dataset is designed to cover a wide range of structured diagrams used in professional settings\. We categorize the 1,449 examples into 6 major domains \(L1\) and 15 sub\-domains \(L2\)\. This section provides a detailed breakdown of each category\.

##### Academic Domain

This domain focuses on scientific research, algorithmic logic, and data representation, which are critical for academic communication\.

- •System Architecture & Flow: This category includes algorithmic logic diagrams, technical architecture blueprints, and model training pipelines\. These diagrams typically involve complex node connections and directed edges representing data or control flow\.
- •Neural Networks: Specific to deep learning, these diagrams feature neurons, layer stacking, convolution kernels, and feature maps\. They often require precise alignment and repetitive structures\.
- •Data Visualization: This includes tensor transformations, experimental histograms, scatter plots, and coordinate axes, emphasizing precise quantitative representation\.

##### Software Domain

This domain covers standardized modeling languages essential in software engineering\.

- •UML Class Diagrams: These are typical three\-tier rectangles showing Class, Property, and Method\. They are fundamental for object\-oriented design and require strict adherence to UML standards\.
- •Sequence Diagrams: These logic diagrams use vertical dashed lines \(lifelines\) and horizontal interaction arrows to depict the sequence of messages between objects\.
- •Database ER Diagrams: Entity\-relationship diagrams that illustrate table structures and primary/foreign key associations, crucial for database design\.

##### Business Domain

This domain emphasizes business planning, corporate operations, and product layout\.

- •Product Architecture: These diagrams illustrate product functional modules, middle\-office components, or platform\-level stacks\.
- •Performance Reports: PPT\-style summaries featuring KPI figures and titled reporting slides, often used in business presentations\.
- •Business Strategy Models: This includes classic management models such as SWOT analysis, PDCA cycles, and Porter’s Five Forces\.

##### Management Domain

This domain is used for organizational collaboration, task tracking, and process standardization\.

- •Gantt Progress: Charts used for project schedule tracking\.
- •Organizational Hierarchy: Diagrams depicting the structure of an organization\.
- •Project Flowcharts: Standard Operating Procedures \(SOPs\) and workflow diagrams\.

##### UIUX Domain

This domain involves interaction design prototypes for digital products\.

- •Web Prototypes: Wireframes and low\-fidelity prototypes for websites\.
- •Mobile Wireframes: Interface designs tailored for mobile devices with specific aspect ratios\.

##### General Domain

This domain includes general logical thinking tools\.

- •Mind Maps: Characterized by radiant or tree\-like logical connection diagrams emanating from a central topic, used for brainstorming and organizing ideas\.

### B\.2Data Source Details

To ensure transparency and reproducibility, we detail our four primary data sources:

1. 1\.Open\-Source Template Repositories \(40%\): Sourced from public GitHub repositories hosting ‘\.drawio‘ or ‘\.xml‘ templates \(e\.g\., ‘jgraph/drawio\-diagrams‘, architecture\-templates\)\.
2. 2\.Open Access Academic Papers \(30%\): Extracted from arXiv sources under CC\-BY licenses, focusing on CS/AI system architectures\.
3. 3\.Anonymized Corporate Diagrams \(20%\): Internal datasets from verified industry partners\. All text entities were anonymized \(e\.g\., "Feature A" instead of specific product names\) and PII was scrubbed using regex and NER pipelines\.
4. 4\.Permissively Licensed Web Diagrams \(10%\): Crawled from technical blogs and documentation sites explicitly marked as CC\-BY or public domain\.

### B\.3Data Synthesis Pipeline

The VCG\-Bench dataset is constructed using a two\-stage synthesis pipeline\. Gemini\-3\-Pro is used to generate intermediate structured captions, while candidatemxGraphXML files are synthesized with multiple VLM candidates and retained through executability, rendering\-consistency, and human\-verification filters\. This data\-construction process is separate from the formal benchmark evaluation, where the evaluated models are compared on the same released Task 1 and Task 2 inputs\.

Stage 1: Image\-to\-mxGraphXML \(Task 1\)

1. 1\.Source Collection: Images gathered from public repositories \(Design2Code, UMLModels\) and web crawls\.
2. 2\.Screening: Filtered for diagrammatic clarity and content\.
3. 3\.Description Generation: Intermediate structured JSON describing ‘components‘, ‘spatial\_layout‘, and ‘overview‘\.
4. 4\.Code Generation \(Task 1\): Converting the description and the original image intomxGraphModelmxGraphXML\.
5. 5\.Verification: Rendering themxGraphXML and filtering based on visual similarity\.

Stage 2: Instruction Synthesis \(Task 2\)

1. 1\.Base Selection: High\-qualitymxGraphXMLs from Stage 1 are selected as ground truth\.
2. 2\.Instruction Generation: Creating instructions with varying difficulty levels \(Easy, Medium, Hard\)\.
3. 3\.Execution & GT Generation: Applying edits via code modification to generate the "After"mxGraphXML state\.
4. 4\.Atomic Operations: Edits are composed of 14 atomic operations \(e\.g\.,add\_node,change\_color,reroute\_edge\)\.

### B\.4Dataset Statistics and Difficulty

Task 1 Difficulty \(mxGraphXML Token Count\):

- •Easy\(< 8,645 tokens\): 33\.0% \(478 images\)
- •Medium\(8,645 \- 14,000 tokens\): 49\.8% \(721 images\)
- •Hard\(\> 14,000 tokens\): 17\.3% \(250 images\)

Statistics: Mean 11,024, Median 10,009\.

Task 2 Difficulty \(Atomic Operations\):

- •Easy: 1\-2 atomic operations\.
- •Medium: 3\-4 atomic operations\.
- •Hard: 5\-7 atomic operations\.

Total: 1,005 editing instructions derived from 335 unique base diagrams\.

### B\.5Dataset and Licensing

This appendix provides additional details on the dataset composition, licensing, and curation policies supporting VCG\-Bench\.

- •Sources: publicly available diagrams from research papers and open repositories\.
- •Licensing: diagrams are included under compliant open licenses; we respect attribution and redistribution terms\.
- •Content: non\-template, compositional layouts featuring mixed shapes, connectors, and textual labels\.
- •Diversity: balanced styles \(palettes, line types, arrowheads\), densities, and connector usage to avoid narrow template bias\.

Sources and Diversity

We specifically select diagrams that emphasize conceptual/system sketches and mixed\-shape schemas over rigid templates\. Diversity targets include:

- •Shape categories: rectangles, rounded boxes, ellipses, diamonds, custom shapes\.
- •Connectors: directed edges with single/multiple waypoints, orthogonal/curved lines, different arrowheads\.
- •Text: node labels, inline annotations, legends, captions\.

## Appendix CAdditional Information

### C\.1Annotation Schema Details

Each image is annotated at element\-level granularity for robust structured encoding\.

- •Nodes: shape type, bounding box, size, style \(fill/line\), text content\.
- •Edges: source/target attachments, waypoints, line style, arrowheads, labels \(if any\)\.
- •Groups/Layers: hierarchical organization to preserve edit semantics and complex layouts\.
- •Relations: attachment, grouping, and layer membership enabling consistent edits \(move/resize without breaking connectors\)\.

### C\.2Ethical Considerations and Release

We comply with licensing, attribution, and privacy norms\. Sensitive or proprietary diagrams are excluded\. Release artifacts include annotated images and valid, editable XML to enable reproducible evaluation\.

### C\.3Discussion

Beyond evaluation, VCG\-Bench serves as a scalable data engine capable of generating high\-quality, verified visual\-code pairs, addressing the chronic scarcity of structured diagrammatic data for VLM training\.

Figure[10](https://arxiv.org/html/2605.15677#A3.F10)illustrates the data flywheel mechanism enabled by our pipeline\. The flywheel operates through a self\-reinforcing cycle with four interconnected components: \(1\)VCG Pipeline: The pipeline processes raw input through three sequential stages: evaluation \(assessment and validation\), filtering \(selection and refinement\), and enhancement \(improvement and augmentation\), ultimately producing high\-quality training data\. The pipeline receives input from the Data Flywheel, creating an iterative improvement mechanism\. \(2\)Data Flywheel: A circular mechanism with four colored segments \(orange, red, teal, green\) arranged in a continuous clockwise cycle, driving the pipeline’s input and receiving feedback to refine the data generation process\. \(3\)Model Training Applications: The high\-quality training data serves three distinct model types:VLM \(Vision2Text\)for vision\-to\-text conversion,LLM \(XML2XML\)for XML\-to\-XML processing, andDiffusion \(Vision2Vision\)for vision\-to\-vision generation\. Each model type benefits from the structured visual\-code pairs with precise annotations\. \(4\)Feedback Loop: Human feedback \(represented by stylized human figures with stars\) flows from the training data back to the Data Flywheel, completing the self\-improving cycle\. This feedback mechanism allows the system to incorporate human judgment and domain expertise, refining both data quality and model performance\. The flywheel addresses a fundamental bottleneck in visual code generation research: the lack of large\-scale, high\-quality training data with precise structural annotations\. By providing both evaluation benchmarks and scalable data generation capabilities with enhanced controllability and explainability, VCG\-Bench contributes to advancing the field beyond mere performance measurement, enabling a continuous improvement cycle for visual code generation systems\. As frontier VLMs depend on efficient long\-context inference to handle the verbose structured outputs inherent tomxGraphXML\(Liuet al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib69)\), the verifiable visual\-code pairs produced by VCG\-Bench offer a targeted training signal for specializing these large models in diagram\-centric tasks\.

![Refer to caption](https://arxiv.org/html/2605.15677v1/figures/flywheel.png)Figure 10:VCG\-Bench Pipeline as a crucial Data Flywheel, delivering high\-quality training data for VLM, LLM, and Diffusion models with enhanced controllability and explainability\.

## Appendix DTechnical Specifications

### D\.1mxGraph XML Schema Overview

Ground\-truthmxGraphXML follows the mxGraph structure and loads without repair in standard editors\. A minimal illustrative snippet:

1<mxGraph\>

2<root\>

3<mxCellid="0"/\>

4<mxCellid="1"parent="0"/\>

5<mxCellid="2"value="Process"style="rounded=1;whiteSpace=wrap;"

6vertex="1"parent="1"\>

7<mxGeometryx="60"y="40"width="120"height="60"as="geometry"/\>

8</mxCell\>

9<mxCellid="3"edge="1"parent="1"source="2"target="4"

10style="endArrow=block;orthogonalLoop=1;"\>

11<mxGeometryrelative="1"as="geometry"\>

12<mxPointx="120"y="100"as="targetPoint"/\>

13</mxGeometry\>

14</mxCell\>

15<mxCellid="4"value="Output"vertex="1"parent="1"\>

16<mxGeometryx="260"y="40"width="120"height="60"as="geometry"/\>

17</mxCell\>

18</root\>

19</mxGraph\>

This structure encodes nodes, connectors, geometry, and styles for accurate reconstruction and editability\.

### D\.2Comparative Analysis of Visual Representation Formats

To contextualize the significance ofmxGraph, we provide a deeper comparison withSVG, HTML, andPowerPoint\(PPT\) on three critical dimensions:

- •Semantic Depth vs\. Graphical Primitives: UnlikeSVG, which primarily defines low\-level vector paths and shapes,mxGraph’smxGraphXML schema explicitly encodes semantic entities \(e\.g\., nodes, actors\) and their logical connections \(e\.g\., edges, directional flows\)\. This abstraction allows models to grasp the underlying system topology rather than just pixel positions\.
- •Structured Editability vs\. Layout Rigidity:HTML/CSSis optimized for document flow and responsive layouts but struggles with the arbitrary spatial relationships inherent in diagrams\. In contrast,mxGraphoffers a flexible canvas that supports precise, absolute positioning while maintaining the structural integrity required for diagrammatic logic, enabling robust round\-trip editing unavailable in rigid DOM structures\.
- •Open Interpretability vs\. Proprietary Formats: WhilePowerPoint\(PPT\) is widely used, its proprietary and complex file structure hinders programmatic generation and fine\-grained manipulation\.mxGraphutilizes an open, human\-readablemxGraphXML format that bridges the gap between visual design and code\-based reasoning, making it an ideal standard for evaluating structured generation capabilities in AI agents\.

Format Specification:mxGraphModelXML \(Draw\.io\)The Draw\.io diagram format is based on anmxGraphModelXML document\. In this benchmark, we assume the following minimal structural requirements for a document to be considered valid:•Document structure: ThemxGraphXML document contains an<mxGraphModel\>element with a<root\>child\. The<root\>element includes \(at minimum\)<mxCell id="0"/\>and<mxCell id="1" parent="0"/\>\.•Cell metadata: Each diagram element is encoded as anmxCellwith attributes includingid\(unique identifier\),parent\(hierarchical containment\),value\(label text, if present\), andstyle\(visual encoding\)\.•Geometric and connectivity information: Node placement is specified via an<mxGeometry\>element withx,y,width, andheightattributes\. Edge connectivity is specified using thesourceandtargetattributes on the correspondingmxCell\.For brevity, we refer to this representation asmxGraphXMLthroughout\.

## Appendix EPrompt Templates and Instructions

We provide the core prompt templates used in our data synthesis and evaluation pipelines\. These prompts utilizegemini\-3\-pro\-previewfor high\-quality generation and reasoning\.

### E\.1Task 1: Diagram Generation Prompts

#### E\.1\.1Flowchart Classification Prompt

Purpose: Determine whether an image is a flowchart/architecture diagram that can be faithfully reconstructed using Draw\.io/XML\.

1AnalyzetheprovidedimageanddetermineifitisaflowchartorarchitecturediagramsuitableforDraw\.io/XMLreconstruction\.

2

3\*\*Criteria\*\*:Theimageshouldcontaindiscretenodes/shapesconnectedbyarrowsorlines,representingasystem,process,orworkflow\.Excludedatavisualizations\(charts,plots\),screenshots,photographs,andimageswithfigurativecharacters\.

4

5\*\*OutputFormat\(JSON\)\*\*:

6\{

7"is\_candidate":true/false,

8"score":0\.0\-1\.0,

9"diagram\_type":"flowchart\|architecture\|network\|dataflow\|other\|null",

10"reason":"Briefexplanation"

11\}

Listing 1:Prompt for Flowchart Classification
#### E\.1\.2Diagram Description Generation Prompt

Purpose: Generate a detailed structured description of the diagram \(JSON format\)\.

1AnalyzetheprovideddiagramimageandgenerateastructuredJSONdescription\.

2

3\*\*Context\*\*:\{context\_text\}\(optional,useifprovided\)

4

5\*\*Requirements\*\*:

6\-Identifyallcomponents\(nodes,shapes,textlabels\)withtheirstylesandpositions

7\-Traceallconnections/arrowswiththeirproperties\(from/to,color,style,routing\)

8\-Describespatiallayoutandrelationships

9\-Extractcolorpaletteandvisualstyles

10

11\*\*OutputFormat\*\*:OutputONLYaJSONobject\(noMarkdown\)withfields:

12\-"image\_type":string

13\-"overview":string

14\-"components":\[\{"name":string,"count":number,"description":string\}\]

15\-"spatial\_layout":\{"primary\_layout":string,"relative\_positions":\[string\],"alignment":\[string\]\}

16\-"flow\_and\_topology":\{"primary\_flow":string,"structures":\[string\]\}

17\-"arrows":\{"summary":\{\.\.\.\},"details":\[\{\.\.\.\}\]\}\(optional\)

18\-"component\_styles":\[\{\.\.\.\}\]\(optional\)

19\-"color\_palette":\{\.\.\.\}\(optional\)

20\-Otheroptionalfieldsasneeded

21

22Usetheimageasgroundtruthforalldescriptions\.

Listing 2:Prompt for Diagram Description Generation
#### E\.1\.3XML Generation Prompt

Purpose: Generate Draw\.iomxGraphXML code based on the image and description\.

1YouareaDraw\.ioXMLgenerationexpert\.Generateacomplete,validDraw\.ioXMLfilebasedonthe\[JSONDescriptionDraft\]and\[OriginalImage\]\.

2

3\*\*Inputs\*\*:

4\-\[JSONDescriptionDraft\]:\{json\_description\}

5\-\[OriginalImage\]:\(useasgroundtruthforverification\)

6

7\*\*XMLRequirements\*\*:

81\.OutputONLYXMLcontent\(noMarkdownfences,noexplanations\)

92\.Startwith‘<mxfile\>‘andendwith‘</mxfile\>‘

103\.‘<mxGraphModel\>‘mustinclude‘grid="1"gridSize="10"guides="1"page="1"‘

114\.Allelementsmustbe‘<mxCell\>‘with:

12\-Unique‘id‘\(globallyunique,addsuffixesifneeded\)

13\-‘value‘,‘style‘,‘parent‘attributes

14\-‘<mxGeometry\.\.\.as="geometry"/\>‘child

155\.Edgecells\(‘edge="1"‘\)mustNOTbeself\-closing;mustcontain‘<mxGeometryrelative="1"as="geometry"/\>‘

166\.XMLescaping:Use‘&lt;‘,‘&gt;‘,‘&amp;‘inattributevalues

177\.Noexternalimages\(no‘shape=image‘withURLs\)

18

19\*\*GenerationRules\*\*:

201\.Use\[OriginalImage\]asgroundtruth;\[JSONDescriptionDraft\]asreference

212\.Layout:Allcoordinatesmustbemultiplesof10\.Centerthediagramoncanvas\(don’tstackintop\-leftcorner\)

223\.Components:Createnodesfrom‘components‘and‘component\_styles‘\.UsestylesfromJSONwhenavailable

234\.Arrows:Createexactlyone‘<mxCelledge="1"\>‘foreachitemin‘arrows\.details‘:

24\-Use‘style‘field\(solid/dashed\)

25\-Set‘startArrow‘/‘endArrow‘basedon‘heads‘field\(doublearrowif‘is\_bidirectional:true‘\)

26\-Use‘routing‘field:add‘edgeStyle=orthogonalEdgeStyle;‘fororthogonal,use‘<Arrayas="points"\>‘forcurved

27\-Set‘exitX‘,‘exitY‘,‘entryX‘,‘entryY‘basedon‘connection\_ports‘\(e\.g\.,"middle\_right"\-\>‘exitX=1;exitY=0\.5;‘\)

285\.Z\-Order:Createbackgroundelements\(with‘is\_background:true‘\)first,thenforegroundelements

296\.Complexstructures:For‘relationships\_extended‘with‘type:"describes\_subgraph"‘or‘"describes\_structure"‘,rebuildthestructureusing‘<mxCell\>‘primitivesbasedonthe‘note‘description

30

31\*\*Validation\*\*:Ensureall‘id‘sareunique,alledgeshavepropersource/targetreferences,andXMLiswell\-formed\.’

Listing 3:Prompt for XML Generation

### E\.2Task 2: Diagram Editing Prompts

#### E\.2\.1Instruction Generation Prompt

Purpose: Generate 3 editing instructions of different difficulty levels \(Easy/Medium/Hard\) for a given diagram\.

1Youareaprofessionaldiagrameditinginstructiongenerationexpert\.Yourtaskistogenerate\*\*3instructionsofdifferentdifficultylevels\*\*forthegivendiagram\.

2

3

4

5\*\*\\texttt\{mxGraph\}XMLStructure\*\*:

6‘‘‘xml

7\{xml\_content\}

8‘‘‘

9

10\*\*RenderedImage\*\*:Pleaseexaminetheprovidedrenderedimagetounderstandthevisualstructureofthediagram\.

11

12\-\-\-

13

14

15

16Thefollowingare\*\*14predefinedatomicoperations\*\*,eachbeingthemostbasic,indivisibleoperationunitindiagramediting:

17

18

191\.\*\*ModifyNodeColor\*\*:Changethefillcolororbordercolorofanode

202\.\*\*ModifyNodeShape\*\*:Changetheshapetypeofanode\(rectangle,circle,diamond,etc\.\)

213\.\*\*ModifyNodeSize\*\*:Changethewidthorheightofanode

224\.\*\*ModifyNodeText\*\*:Changethetextcontentdisplayedinsideanode

23

24

255\.\*\*DeleteNode\*\*:Deleteanode\(withouthandlingconnections\)

266\.\*\*AddNode\*\*:Addanewnodeataspecifiedlocation

277\.\*\*MoveNode\*\*:Changethepositionofanode

28

29

308\.\*\*ModifyConnectionColor\*\*:Changethecolorofaconnectionline

319\.\*\*ModifyConnectionStyle\*\*:Changethestyleofaconnectionline\(solid,dashed,thickness,etc\.\)

3210\.\*\*ModifyConnectionArrow\*\*:Changethearrowstyleofaconnectionline

33

34

3511\.\*\*DeleteConnection\*\*:Deleteaconnectionline

3612\.\*\*AddConnection\*\*:Addaconnectionlinebetweentwonodes

3713\.\*\*RedirectConnection\*\*:Changethestartorendpointofaconnectionline

3814\.\*\*UpdateConnectionPath\*\*:Updatetheconnectionlinepathwhennodesmove

39

40\-\-\-

41

42

43

44Youneedtogenerate3instructionsofdifferentdifficultylevels,wheredifficultyis\*\*completelydeterminedbythenumberofatomicoperations\*\*:

45

46

47\-\*\*Requirement\*\*:Theinstructionmustcontain\*\*1\-2\*\*oftheabove14atomicoperations

48

49

50\-\*\*Requirement\*\*:Theinstructionmustcontain\*\*3\-4\*\*oftheabove14atomicoperations

51

52

53\-\*\*Requirement\*\*:Theinstructionmustcontain\*\*5\-7\*\*oftheabove14atomicoperations

54

55\-\-\-

56

57

58

59\*\*MostImportantPrinciple\*\*:Instructionsmustbebasedonthe\*\*renderedimage\*\*thatuserssee,usingnaturallanguage,notbasedon\\texttt\{mxGraph\}XMLcodeortechnicalidentifiers\.

60

61

62

63\*\*HighestPriority\*\*:Directlyusethetextcontentdisplayedonnodes

64

651\.\*\*NodeTextContentDescription\(Mostnaturalandaccurate\)\*\*:

66\-"Changethe’Start’nodetored"

67\-"Deletethe’DataProcessing’node"

68

692\.\*\*SemanticDescription\*\*\(whennodesdon’thavecleartextbutsemanticsareclear\):

70\-"Startnode","Endnode"

71

723\.\*\*PositionDescription\*\*\(whennodeshavenotextortextisunclear\):

73\-"Thetopmostnode","Theleftmostrectangle","Themiddlecircle"

74

754\.\*\*VisualFeatureDescription\*\*\(assupplementarypositioning\):

76\-"Thelargestrectangle","Therednode"

77

785\.\*\*RelativePositionDescription\*\*:

79\-"Thearrowbetween’Start’and’End’"

80

81\#\#\#\[PROHIBITED\]StrictlyProhibitedDescriptionMethods:

82

831\.\*\*TechnicalIdentifiers\*\*\(Prohibited\):

84\-\[X\]"Nodewithid’node\_1’"

85\-\[X\]"edge\_5"

86

872\.\*\*CoordinateDescription\*\*\(Prohibited\):

88\-\[X\]"Nodeatcoordinates\(200,300\)"

89

903\.\*\*XMLAttributeReference\*\*\(Prohibited\):

91\-\[X\]"Nodewithshape=’rectangle’"

92

93\-\-\-

94

95\#\#OutputFormat\(StrictJSON\)

96

97PleaseoutputaJSONobjectcontaining3instructionsinthefollowingformat:

98

99‘‘‘json

100\{

101"instructions":\[

102\{

103"difficulty":"Easy",

104"instruction":"Changethe’Start’nodetored",

105"atomic\_operations":\[

106\{

107"operation\_id":1,

108"operation\_type":"ModifyNodeColor",

109"description":"Changethecolorofthe’Start’nodetored"

110\}

111\],

112"atomic\_operation\_count":1,

113"target\_elements":\[

114\{

115"type":"node",

116"visual\_description":"The’Start’node"

117\}

118\],

119"executable":true

120\},

121\{

122"difficulty":"Medium",

123\.\.\.

124\},

125\{

126"difficulty":"Hard",

127\.\.\.

128\}

129\]

130\}

131‘‘‘’

Listing 4:Prompt for Instruction Generation
#### E\.2\.2Instruction Decomposition Prompt \(XDRFR\)

Purpose: Decompose a natural language editing instruction into a series of verifiable Yes/No questions\.

1DecomposethefollowingdiagrammodificationinstructionintoaseriesofYes/Noquestions\.Thesequestionswillbeusedtoverifywhethertheinstructionhasbeencorrectlyfollowed\.

2Thesequestionswillbeverifiedbasedon\\texttt\{mxGraph\}XMLcode,notrenderedimages\.

3

4\*\*ImportantNote\*\*:Theinstructionusesnaturallanguagedescriptions\(basedonrenderedimages\),butquestionsneedtobeconvertedto\\texttt\{mxGraph\}XMLcode\-levelverification\.

5Forexample,iftheinstructionsays"Changethe’Start’nodetored",thequestionshouldbeconvertedto"HasthefillColorattributeofthenodewithvalue’Start’beenchangedtored?"

6

7Instruction:"\{instruction\}"

8

9\*\*Requirements:\*\*

10

11\*\*1\.Targetonlytheinstructionitself\*\*

12\-Carefullyanalyzethespecificmodificationcontentmentionedintheinstruction

13\-DecomposeeachspecificrequirementintheinstructionintoindependentYes/Noquestions

14\-Donotaddcontentorchecksnotmentionedintheinstruction

15

16\*\*2\.Questionformatrequirements\*\*

17\-Eachquestionshouldbeanswerablewith"Yes"or"No"

18\-Questionsshouldbeverifiablebycomparingtheoriginal\\texttt\{mxGraph\}XMLcodeandthemodified\\texttt\{mxGraph\}XMLcode

19\-Questionsshouldbeclearandunambiguous

20\-Questionsshoulddirectlycorrespondtothespecificmodificationrequirementsintheinstruction

21\-Questionsshouldfocuson\\texttt\{mxGraph\}XMLcode\-levelmodifications\(attributevalues,elementexistence,structuralchanges,etc\.\)

22\-\*\*Key\*\*:Convertnaturallanguagedescriptionsto\\texttt\{mxGraph\}XMLcode\-levelverificationmethods

23

24\*\*3\.Naturallanguageto\\texttt\{mxGraph\}XMLmappingrules\*\*

25\-\*\*Nodetextdescription\*\*\(e\.g\.,"Changethe’Start’node\.\.\."\)\-\>Findnodesin\\texttt\{mxGraph\}XMLwherethe‘value‘attributecontainsthattext

26\-\*\*Positiondescription\*\*\(e\.g\.,"thetopmostnode"\)\-\>Needtocombine‘mxGeometry‘coordinateinformationtolocate\(smallestycoordinate\)

27\-\*\*Semanticdescription\*\*\(e\.g\.,"startnode"\)\-\>Findnodeswherethe‘value‘attributecontainsrelatedsemantics

28\-\*\*Colordescription\*\*\(e\.g\.,"changetored"\)\-\>Verify‘fillColor‘or‘strokeColor‘inthe‘style‘attribute\(supportcolornamesorhexcodes\)

29\-\*\*Shapedescription\*\*\(e\.g\.,"changetocircle"\)\-\>Verify‘shape‘orrelatedattributesinthe‘style‘attribute

30\-\*\*Connectiondescription\*\*\(e\.g\.,"arrowconnecting’Input’and’Output’"\)\-\>Findnodeswhere‘source‘and‘target‘attributescorrespond\(matchbyvalueattribute\)

31

32\*\*OutputFormat\(JSON\):\*\*

33\{

34"decomposed\_questions":\[

35"HasthefillColorattributeofthenodewithvalue’Start’beenchangedtored\(\#FF0000orred\)?",

36"Doesthenodewithvalue’Start’stillexistinthemodified\\texttt\{mxGraph\}XMLcode?"

37\]

38\}

Listing 5:Prompt for XDRFR Instruction Decomposition
#### E\.2\.3XML Editing Prompt

Purpose: ModifymxGraphXML code according to the instruction \(incremental format\)\.

1YouareaprofessionalDraw\.ioXMLeditingexpert\.YourtaskistoeditthegivenXMLaccordingtothemodificationinstruction\.

2

3

4‘‘‘xml

5\{original\_xml\}

6‘‘‘

7

8

9\{instruction\}

10

11

121\.\*\*Onlyoutputmodifiedparts\*\*:Tosavetokens,youonlyneedtooutputthemodified\\texttt\{mxGraph\}XMLfragments,notthecomplete\\texttt\{mxGraph\}XML

132\.\*\*Incrementalformat\*\*:OutputJSONformatcontainingallplacesthatneedmodification\(becausetheremaybemultipleplacestomodify\)

143\.\*\*Formatrequirements\*\*:

15\-Eachmodificationcontainstwofields:

16\-‘original\_fragment‘:Original\\texttt\{mxGraph\}XMLfragment\(theparttobereplaced\)

17\-‘modified\_fragment‘:Modified\\texttt\{mxGraph\}XMLfragment\(thereplacementcontent\)

18\-Iftherearemultiplemodifications,outputmultiplemodificationobjects

194\.\*\*Fragmentrequirements\*\*:

20\-‘original\_fragment‘mustbeacompletefragmentfromtheoriginal\\texttt\{mxGraph\}XML\(canbeacompleteelement,suchas‘<mxCell\>\.\.\.</mxCell\>‘\)

21\-‘modified\_fragment‘isthecorrespondingmodifiedfragment

22\-Fragmentsmustbepreciseenoughtouniquelymatchthepositionintheoriginal\\texttt\{mxGraph\}XML

235\.\*\*Maintaincorrect\\texttt\{mxGraph\}XMLformat\*\*:Modifiedfragmentsmustmaintaincorrectandparseable\\texttt\{mxGraph\}XMLformat

24

25

26

27PleaseoutputaJSONobjectinthefollowingformat:

28

29‘‘‘json

30\{

31"changes":\[

32\{

33"original\_fragment":"<mxCellid=’node\_1’value=’OldText’\.\.\.\>\.\.\.</mxCell\>",

34"modified\_fragment":"<mxCellid=’node\_1’value=’NewText’\.\.\.\>\.\.\.</mxCell\>"

35\},

36\{

37"original\_fragment":"<mxCellid=’edge\_1’\.\.\.\>\.\.\.</mxCell\>",

38"modified\_fragment":"<mxCellid=’edge\_1’\.\.\.\>\.\.\.</mxCell\>"

39\}

40\]

41\}

42‘‘‘

43

44\*\*ImportantNotes\*\*:

45\-Iftheinstructioninvolvesonlyonemodification,the‘changes‘arraycontainsonlyoneelement

46\-Iftheinstructioninvolvesmultiplemodifications\(e\.g\.,"deletenodeAandnodeB"\),the‘changes‘arraycontainsmultipleelements

47\-‘original\_fragment‘mustbeacompleteandprecisefragmentfromtheoriginal\\texttt\{mxGraph\}XML

48\-Fordeletionoperations,‘modified\_fragment‘isanemptystring‘""‘

49\-OnlyoutputJSON,donotincludeanyexplanatorytextorMarkdowncodeblockmarkers

Listing 6:Prompt for XML Editing

### E\.3Evaluation\-Related Prompts

#### E\.3\.1SCS \(Style Consistency Score\)

##### SCS Prompt \(Task 1\)

Purpose: Evaluate style consistency between the generated diagram and the original diagram \(Task 1\)\.

1Youareaprofessionaldiagramdesignreviewer\.Pleasecomparethe"originaldiagram"and"generateddiagram"andconductasystematicevaluationfollowingthesesteps:

2

3Originaldiagram:

4\[Image:\{original\_path\}\]

5

6Generateddiagram:

7\[Image:\{generated\_path\}\]

8

9\*\*EvaluationSteps:\*\*

10

11\*\*Step1:AttributeExtraction\*\*

12Pleasecarefullyanalyzetheoriginaldiagramandextractthefollowingkeyattributes:

13\-Colorscheme:Listthemaincolors’hexadecimalvaluesorcolorfamilynames\(e\.g\.,"bluefamily","warmtones"\)

14\-Fontstyle:Fontweight\(thin/medium/bold\),size\(small/medium/large\),fontfamily\(ifany\)

15\-Layoutstructure:Overallstructuretype\(tree/ring/linear/grid/freelayout\)

16\-Visualelements:Nodeshapes\(circle/rectangle/diamond,etc\.\),linestyles\(solid/dashed/thickness\)

17

18\*\*Step2:DifferenceComparison\*\*

19Checkthedeviationoftheaboveattributesinthegenerateddiagramonebyone:

20\-Isthecolorschemeconsistent?Iftherearedifferences,describethespecificdifferences

21\-Doesthefontstylematch?

22\-Isthelayoutstructurethesame?Areelementpositionsandspacingrelationshipsconsistent?

23\-Arevisualelementstylesconsistent?

24

25\*\*Step3:DimensionScoring\(0\-10scale\)\*\*

26Pleasescorethefollowingthreedimensionsseparately\(0\-10points,keeponedecimalplace\):

271\.Visualstyleconsistency\(color,linethickness,nodestyle\):\_\_\_

282\.Layoutstructureconsistency\(elementpositions,spacing,spatialrelationships\):\_\_\_

293\.Aestheticquality\(alignment,visualbalance,overallbeauty\):\_\_\_

30

31\*\*Step4:FinalScoreCalculation\*\*

32\-Calculatetheaverageofthethreedimensionscores

33\-Dividetheaverageby10tonormalizeto0\-1

34\-Example:Dimensionscores\[8\.5,7\.0,9\.0\],average=8\.17,finalscore=0\.817

35

36\*\*OutputFormat\(JSON\):\*\*

37\{

38"analysis":\{

39"original\_attributes":\{

40"color\_scheme":"\.\.\.",

41"font\_style":"\.\.\.",

42"layout\_structure":"\.\.\.",

43"visual\_elements":"\.\.\."

44\},

45"differences":\[

46"Differencedescription1",

47"Differencedescription2"

48\],

49"dimension\_scores":\{

50"visual\_style\_consistency":8\.5,

51"layout\_consistency":7\.0,

52"aesthetic\_quality":9\.0

53\},

54"average\_score":8\.17

55\},

56"final\_score":0\.817

57\}

58

59\*\*ImportantNotes:\*\*

60\-Pleasestrictlyfollowthescoringcriteriaandavoiddirectlygivinghighscores\(0\.85\+\)

61\-Youmustcompleteattributeextractionanddifferencecomparisonbeforegivingdimensionscores

62\-Ifthegenerateddiagramhasobviousvisualerrors\(suchaselementoverlap,textmisalignment\),pointsshouldbedeductedinthecorrespondingdimension

63\-Thefinalscoremustbecalculatedbasedontheaverageofdimensionscores,notestimateddirectly’

Listing 7:SCS Prompt for Task 1
##### SCS Prompt \(Task 2\)

Purpose: Evaluate style and aesthetic consistency between the modified diagram and the original diagram \(Task 2; only style and aesthetics are evaluated\)\.

1Youareaprofessionaldiagramdesignreviewer\.Pleasecomparethe"OriginalDiagram\(Geminigenerated\)"andthe"ModifiedDiagram\(Modelgenerated\)"toevaluatewhetherthemodifieddiagrammaintainstheoriginal\*\*styleandaesthetic\*\*\.

2

3\*\*ImportantNotes\*\*:

4\-Themodifieddiagramwilldefinitelyhavechanges\(elementsmaybedeleted,added,modified\),whichis\*\*normal\*\*andshouldnotresultinscorededuction

5\-\*\*Onlyevaluatestyleconsistencyandaestheticconsistency\*\*,donotevaluateelementcompleteness,textconsistency,orothercontentconsistency

6\-Wefocuson:whetherthemodifieddiagramoverallstillconformstotheoriginaldiagram’sstyleandaestheticstandards

7

8OriginalDiagram\(Geminigenerated\):

9\[Image:\{gemini\_rendered\_path\}\]

10

11ModifiedDiagram\(Modelgenerated\):

12\[Image:\{modified\_rendered\_path\}\]

13

14\*\*EvaluationSteps:\*\*

15

16\*\*Step1:StyleAttributeExtraction\*\*

17Pleasecarefullyanalyzetheoriginaldiagramandextractthefollowing\*\*styleandaesthetic\*\*relatedkeyattributes:

18\-\*\*ColorScheme\*\*:Maincolorfamilies,tones,saturationcharacteristics\(e\.g\.,"bluefamily","warmtones","highsaturation"\)

19\-\*\*FontStyle\*\*:Overallstylecharacteristicsoffontweight,size,fontfamily

20\-\*\*VisualElementStyle\*\*:Uniformityofnodeshapes,uniformityoflinestyles,overallvisualstyle

21\-\*\*AestheticFeatures\*\*:Overalllayoutbalance,alignmentmethods,visualhierarchy,spaceutilization

22

23\*\*Step2:StyleandAestheticConsistencyEvaluation\*\*

24Evaluatethemodifieddiagram’sperformanceinthefollowingaspects:

25\-\*\*ColorStyleConsistency\*\*:Dounmodifiedpartsmaintaintheoriginalcolorstyle?Istheoverallcolorschemeharmonious?

26\-\*\*VisualStyleConsistency\*\*:Dounmodifiedpartsmaintaintheoriginalvisualstyle\(nodeshapes,linestyles,etc\.\)?

27\-\*\*AestheticQuality\*\*:Isthemodifieddiagramoverallstillbeautiful,balanced,andprofessional?Arethereobviousvisualerrors\(elementoverlaps,textmisalignment,layoutchaos\)?

28

29\*\*ImportantEvaluationPrinciples\*\*:

30\-\[YES\]\*\*Shouldevaluate\*\*:Styleconsistencyofunmodifiedparts,overallaestheticquality,visualerrors

31\-\[NO\]\*\*ShouldNOTevaluate\*\*:Whetherelementcountsareconsistent,whethertextcontentisconsistent,whetherelementsarecomplete\(thesechangesarenormal\)

32

33\*\*Step3:DimensionScoring\(0\-10scale\)\*\*

34Pleasescorethefollowingtwodimensionsseparately\(0\-10points,keeponedecimalplace\):

351\.\*\*StyleConsistency\*\*\(colorstyle,visualelementstyle,overallstylecharacteristics\):\_\_\_

36\-Evaluatewhetherunmodifiedpartsmaintaintheoriginalstylecharacteristics

37\-Evaluatewhethertheoverallstyleiscoordinatedandunified

382\.\*\*AestheticQuality\*\*\(visualbalance,alignment,overallbeauty,presenceofobviousvisualerrors\):\_\_\_

39\-Evaluatewhetherthemodifieddiagramisstillbeautifulandprofessional

40\-Ifthereareobviousvisualerrors\(elementoverlaps,textmisalignment,layoutchaos\),deductpoints

41

42\*\*Step4:FinalScoreCalculation\*\*

43\-Calculatetheaverageofthetwodimensionscores

44\-Dividetheaverageby10tonormalizeto0\-1

45\-Example:Dimensionscores\[8\.5,9\.0\],average=8\.75,finalscore=0\.875

46

47\*\*OutputFormat\(JSON\):\*\*

48\{

49"analysis":\{

50"original\_attributes":\{

51"color\_scheme":"\.\.\.",

52"font\_style":"\.\.\.",

53"visual\_element\_style":"\.\.\.",

54"aesthetic\_features":"\.\.\."

55\},

56"differences":\[

57"Differencedescription1",

58"Differencedescription2"

59\],

60"dimension\_scores":\{

61"style\_consistency":8\.5,

62"aesthetic\_quality":9\.0

63\},

64"average\_score":8\.75

65\},

66"final\_score":0\.875

67\}

68

69\*\*ImportantNotes:\*\*

70\-Pleasestrictlyfollowthescoringcriteriaandavoiddirectlygivinghighscores\(0\.85\+\)

71\-Youmustcompleteattributeextractionanddifferencecomparisonbeforegivingdimensionscores

72\-Ifthemodifieddiagramhasobviousvisualerrors\(suchaselementoverlap,textmisalignment\),pointsshouldbedeductedintheaestheticqualitydimension

73\-Thefinalscoremustbecalculatedbasedontheaverageofdimensionscores,notestimateddirectly

Listing 8:SCS Prompt for Task 2

#### E\.3\.2CodeVQA

##### CodeVQA Evaluation Prompt

Purpose: Answer CodeVQA questions based onmxGraphXML code\.

1Youareaprofessionaldiagramanalysisexpert\.PleasecarefullyanalyzethefollowingDraw\.io\\texttt\{mxGraph\}XMLcodeandanswerALLthequestionsbelowbasedonthe\\texttt\{mxGraph\}XMLcontent\.

2

3\*\*Generated\\texttt\{mxGraph\}XMLCode:\*\*

4‘‘‘xml

5\{generated\_xml\}

6‘‘‘

7

8\*\*Questions:\*\*

9\{questions\_text\}

10

11\*\*Instructions:\*\*

12\-Analyzethe\\texttt\{mxGraph\}XMLcodestructure,elements,attributes,andcontent

13\-Forcountingquestions:Countspecificelementsorattributesinthe\\texttt\{mxGraph\}XML

14\-Foridentificationquestions:Findspecifictextvalues,IDs,orattributesinthe\\texttt\{mxGraph\}XML

15\-Forrelationshipquestions:Analyzeconnections,hierarchy,orrelationshipsbetweenelementsinthe\\texttt\{mxGraph\}XML

16\-Answereachquestionbasedsolelyonthe\\texttt\{mxGraph\}XMLcodeprovided

17\-Donotmakeassumptionsbeyondwhatisinthe\\texttt\{mxGraph\}XML

18

19\*\*OutputFormat\(JSON\):\*\*

20PleaseprovideanswersinthefollowingJSONformat\(oneanswerperquestion\):

21\{

22"Q1":"<answertoquestion1\>",

23"Q2":"<answertoquestion2\>",

24\.\.\.

25\}

26

27\*\*Important:\*\*

28\-Answereachquestiondirectlywithoutadditionalexplanations

29\-Usetheexactquestionnumbers\(Q1,Q2,Q3,etc\.\)askeys

30\-Provideconciseanswersonly

31\-Baseyouranswersstrictlyonthe\\texttt\{mxGraph\}XMLcodecontent

Listing 9:CodeVQA Evaluation Prompt
##### CodeXQA Question\-Answer Generation Prompt

Purpose: Generate counting, identification, and relationship QA pairs for the CodeXQA benchmark based on the original image\.

1Pleaseanalyzetheprovideddiagramimageandgenerate3question\-answerpairsfortestingsemanticinformationretention\.

2Mustincludethefollowingthreetypes,oneofeach:

31\.Counting:Countthenumberofelementsinthediagram

42\.Identification:Identifyattributesorlabelsofspecificelements

53\.Relationship:Identifyrelationshipsorconnectionsbetweenelements

6

7Image:\[ImagePlaceholder\]

8

9\*\*Requirements\(EnhancedVersion\):\*\*

10

11\*\*1\.DepthRequirements\(ImproveDiscrimination\)\*\*

12\-\*\*Countingquestions\*\*:Shouldnotonlyask"howmanynodes"whichistoosimple\.Shouldincludecountingofspecificattributes,forexample:

13\-"Howmanybluecircularnodesareinthediagram?"

14\-"Howmanydashedconnectionsarethere?"

15\-"Howmanynodeshavelabelsstartingwiththeletter’A’?"

16\-\*\*Identificationquestions\*\*:Shouldtestspecificvisualdetails,forexample:

17\-"Whatisthefillcolorofthenodewithid’node\_5’?"

18\-"Whatisthelabelofthenodelocatedatthetop\-leftcornerofthediagram?"

19\-"WhatcoloristhelineconnectingnodeAandnodeB?"

20\-\*\*Relationshipquestions\*\*:Shouldtestmulti\-levelconnectionsorcomplexrelationships,forexample:

21\-"ThroughwhichintermediatenodescannodeAreachnodeD?"

22\-"WhichnodesareconnectedtobothnodeBandnodeC?"

23\-"Howmanynodesareinthelongestpathfromtherootnodetoaleafnode?"

24

25\*\*2\.SemanticUniqueness\*\*

26\-Ensureeachquestion’sanswerisuniqueandunambiguousinthediagram

27\-Avoidquestionsthatcanbeansweredthroughcommonsenseorcontext

28\-Questionsshouldrequirethemodeltocarefullyobservetheimagetoanswer

29

30\*\*3\.VisualAnchors\*\*

31\-Guidequestionstofocusonkeydetailsandspecificlocationsinthediagram

32\-Usespecificidentifiers\(suchasnodeIDs,positiondescriptions\)toanchorquestions

33\-Preventmodelsfromansweringthroughgeneralreasoningratherthanvisualobservation

34

35\*\*4\.Verifiability\*\*

36\-Answersshouldbespecificandverifiable\(numbers,colorvalues,labeltext,etc\.\)

37\-Avoidsubjectivejudgmentquestions

38

39\*\*OutputFormat\(JSON\):\*\*

40\{

41"qa\_pairs":\[

42\{

43"question":"Howmanybluecircularnodesareinthediagram?",

44"answer":"3",

45"question\_type":"counting",

46"visual\_anchor":"bluecircularnodes"

47\},

48\{

49"question":"Whatisthefillcolor\(inhexadecimal\)ofthenodewithid’anchor’?",

50"answer":"\#FF5733",

51"question\_type":"identification",

52"visual\_anchor":"nodeID’anchor’"

53\},

54\{

55"question":"ThroughwhichintermediatenodescannodeAreachnodeD?Pleaselistinorder\.",

56"answer":"NodeBandNodeC",

57"question\_type":"relationship",

58"visual\_anchor":"pathfromnodeAtonodeD"

59\}

60\]

61\}’

Listing 10:Prompt for CodeXQA QA Generation

#### E\.3\.3XDRFR

##### XDRFR Single Question Evaluation Prompt

Purpose: Evaluate a single question based on the model’s incremental modification output \(JSON\)\.

1Youareaprofessional\\texttt\{mxGraph\}XMLmodificationevaluationexpert\.Pleaseanswerthefollowingquestionwith"Yes"or"No"basedontheprovidedmodeloutputJSON\(containingincrementalmodifications\)andmodificationinstruction\.

2

3\*\*ImportantNote\*\*:Themodificationinstructionusesnaturallanguagedescriptions\(basedonrenderedimages\)\.ThemodeloutputJSONcontainsincrementalmodificationsintheformatoforiginal\_fragmentandmodified\_fragmentpairs\.Youneedtounderstandthenaturallanguageinstruction,thenverifywhetherthemodificationsintheJSONsatisfytheinstructionrequirements\.

4

5\*\*ModelOutputJSON\(IncrementalModifications\):\*\*

6‘‘‘json

7\{model\_output\_json\_str\}

8‘‘‘

9

10\*\*ModificationInstruction\(NaturalLanguage\):\*\*

11"\{instruction\}"

12

13\*\*Question:\*\*

14\{question\}

15

16\*\*Requirements:\*\*

17\-Answerwithonly"Yes"or"No",donotoutputanyothercontent

18\-Donotoutputreasons,explanations,evidence,orjudgmentlogic

19\-AnswerdirectlywithYes/No

20\-MustbasejudgmentonthemodeloutputJSONcontent\(incrementalmodifications\),andverifywhetherthemodificationssatisfytheinstructionrequirements,notspeculation

21

22\*\*OutputFormat\(JSON\):\*\*

23\{

24"answer":"Yes"or"No"

25\}

26

27\*\*JSONFormat:\*\*

28Eachmodificationobjectcontains‘original\_fragment‘\(original\\texttt\{mxGraph\}XML\)and‘modified\_fragment‘\(modified\\texttt\{mxGraph\}XML\)\.

29

30\*\*EvaluationProcess:\*\*

311\.Analyzeeachmodificationbycomparing‘original\_fragment‘and‘modified\_fragment‘

322\.Mapinstructionrequirementstomodifications

333\.Verifycompletenessandcorrectness

34

35\*\*VerificationRules:\*\*

36\-\*\*Textchanges\*\*:Check‘value‘attributein‘modified\_fragment‘

37\-\*\*Colorchanges\*\*:Check‘fillColor‘/‘strokeColor‘in‘style‘attribute\(hexcodeslike

38\-\*\*Position/Size\*\*:Check‘mxGeometry‘coordinates/dimensions

39\-\*\*Add/Remove\*\*:Checkfornewelementsin‘modified\_fragment‘orempty‘modified\_fragment‘forremovals

40\-\*\*Attributes\*\*:Compareattributesbetweenfragments

41

42\*\*Rules:\*\*

43\-Analyzeallmodificationsandverifytheymatchtheinstruction

44\-Verifycompleteness\(allinstructionaspectsaddressed\)andcorrectness\(changesalignwithinstruction\)

45\-Ifuncertain,answer"No"\(conservativestrategy\)

Listing 11:XDRFR Single Question Evaluation Prompt
##### XDRFR Batch Evaluation Prompt

Purpose: Batch evaluate multiple questions based on the model’s output incremental modification JSON\.

1YouareaprofessionalXMLmodificationevaluationexpert\.Pleaseanswerallthefollowingquestionswith"Yes"or"No"basedontheprovidedmodeloutputJSON\(containingincrementalmodifications\)andmodificationinstruction\.

2

3\*\*ImportantNote\*\*:Themodificationinstructionusesnaturallanguagedescriptions\(basedonrenderedimages\)\.ThemodeloutputJSONcontainsincrementalmodificationsintheformatoforiginal\_fragmentandmodified\_fragmentpairs\.Youneedtounderstandthenaturallanguageinstruction,thenverifywhetherthemodificationsintheJSONsatisfytheinstructionrequirements\.

4

5\*\*ModelOutputJSON\(IncrementalModifications\):\*\*

6‘‘‘json

7\{model\_output\_json\_str\}

8‘‘‘

9

10\*\*ModificationInstruction\(NaturalLanguage\):\*\*

11"\{instruction\}"

12

13\*\*QuestionList:\*\*

14\{questions\_text\}

15

16\*\*Requirements:\*\*

17\-Answereachquestionwithonly"Yes"or"No",donotoutputanyothercontent

18\-Donotoutputreasons,explanations,evidence,orjudgmentlogic

19\-AnswereachquestiondirectlywithYes/No

20\-MustbasejudgmentonthemodeloutputJSONcontent\(incrementalmodifications\),andverifywhetherthemodificationssatisfytheinstructionrequirements,notspeculation

21

22\*\*OutputFormat\(JSON\):\*\*

23\{

24"answers":\[

25\{"question":"Completetextofquestion1","answer":"Yes"\},

26\{"question":"Completetextofquestion2","answer":"No"\},

27\.\.\.

28\]

29\}

30

31\*\*JSONFormat:\*\*

32Eachmodificationobjectcontains‘original\_fragment‘\(original\\texttt\{mxGraph\}XML\)and‘modified\_fragment‘\(modified\\texttt\{mxGraph\}XML\)\.

33

34\*\*EvaluationProcess:\*\*

351\.Analyzeeachmodificationbycomparing‘original\_fragment‘and‘modified\_fragment‘

362\.Mapinstructionrequirementstomodifications

373\.Verifycompletenessandcorrectness

38

39\*\*VerificationRules:\*\*

40\-\*\*Textchanges\*\*:Check‘value‘attributein‘modified\_fragment‘

41\-\*\*Colorchanges\*\*:Check‘fillColor‘/‘strokeColor‘in‘style‘attribute\(hexcodeslike

42\-\*\*Position/Size\*\*:Check‘mxGeometry‘coordinates/dimensions

43\-\*\*Add/Remove\*\*:Checkfornewelementsin‘modified\_fragment‘orempty‘modified\_fragment‘forremovals

44\-\*\*Attributes\*\*:Compareattributesbetweenfragments

45

46\*\*Rules:\*\*

47\-Analyzeallmodificationsandverifytheymatchtheinstruction

48\-Verifycompleteness\(allinstructionaspectsaddressed\)andcorrectness\(changesalignwithinstruction\)

49\-Ifuncertain,answer"No"\(conservativestrategy\)

Listing 12:XDRFR Batch Evaluation Prompt

### E\.4Instruction Templates

Expert\-authored directives cover structural, stylistic, textual, and compositional edits:

- •Structural: add/delete nodes or edges; re\-parenting; grid/flow re\-layout\.
- •Stylistic: palette changes; line/fill styles; font family/size/weight updates\.
- •Textual: rename labels; insert annotations; consolidate legends\.
- •Compositional: merge diagrams; extract subgraphs; duplicate and refactor patterns\.

## Appendix FEvaluation Methods and Metrics

Motivated by findings that LLMs degrade substantially when evaluated on*generating*coherent output rather than retrieving facts\(Liuet al\.,[2024a](https://arxiv.org/html/2605.15677#bib.bib66)\), that single\-metric evaluations miss multi\-dimensional capability gaps\(Donget al\.,[2025](https://arxiv.org/html/2605.15677#bib.bib67)\), and that code generation is among the most degradation\-prone fundamental abilities under resource\-constrained deployment\(Liuet al\.,[2026](https://arxiv.org/html/2605.15677#bib.bib68)\), we design a decomposed metric suite \(CodeXQA, XDRFR, ESR, SCS\) that jointly assesses structural accuracy, instruction compliance, and visual fidelity of generatedmxGraphXML\.

### F\.1Question Set Construction

#### F\.1\.1CodeXQA Question Types

CodeXQA evaluates semantic retention through three distinct question types\. Each sample in the benchmark contains exactly one question of each type\.

- •Counting: Evaluates the model’s ability to aggregate statistics about specific visual elements\. - –Example: "How many blue circular nodes are in the diagram?" - –Focus: Color, shape, and object detection\.
- •Identification: Tests the precise retrieval of attributes for specific elements\. - –Example: "What is the fill color \(in hex\) of the node with id ’node\_5’?" - –Focus: Fine\-grained attribute recognition and text\-to\-object mapping\.
- •Relationship: Assesses the understanding of topological connections and flows\. - –Example: "Through which intermediate nodes can node A reach node D?" - –Focus: Graph topology, pathfinding, and connectivity\.

#### F\.1\.2XDRFR Decomposition Logic

For Task 2 \(Instruction Editing\), we employ a “Decompose\-and\-Verify” strategy\. Natural language instructions are broken down into atomic, verifiable checkpoints\.

NL tomxGraphXML Mapping Rules:

- •Text References: "The ’Start’ node"→\\rightarrowmxGraphXML node withvalue="Start"\.
- •Spatial References: "The top\-most node"→\\rightarrowmxGraphXML node withmin⁡\(y\)\\min\(y\)inmxGeometry\.
- •Style References: "Red node"→\\rightarrowmxGraphXML node withfillColor="\#FF0000"instyle\.
- •Connection References: "Link from A to B"→\\rightarrowmxCellwheresourcematches A’s ID andtargetmatches B’s ID\.

Complexity\-Adaptive Question Count:

- •Simple Instructions\(1\-2 atomic ops\): Decomposed into 1\-2 questions\.
- •Medium Instructions\(3\-4 atomic ops\): Decomposed into 2\-3 questions\.
- •Complex Instructions\(5\-7 atomic ops\): Decomposed into 3\-5 questions\.

### F\.2Evaluation Metrics Details

We provide precise mathematical definitions for the key metrics used in VCG\-Bench\.

#### F\.2\.1Task 1: Generation Metrics

1\. Execution Success Rate \(ESR\)ESR measures the proportion of generatedmxGraphXML codes that are syntactically valid and renderable\.

ESR=\{1\.0ifmxGraphXML\_valid∧Render\_success0\.0otherwise\\mathrm\{ESR\}=\\begin\{cases\}1\.0&\\text\{if \}\\text\{\{mxGraph\} XML\\\_valid\}\\land\\text\{Render\\\_success\}\\\\ 0\.0&\\text\{otherwise\}\\end\{cases\}
2\.mxGraphXML Token Count \(XTC\)XTC is a proxy for structural complexity, calculated using the ‘cl100k\_base‘ tokenizer\. Primarily used for image difficulty classification, not as a model performance evaluation metric\.

XTC=Tokenize​\(mxGraphXML\_code,cl100k\_base\)\\mathrm\{XTC\}=\\mathrm\{Tokenize\}\(\\text\{\{mxGraph\} XML\\\_code\},\\text\{cl100k\\\_base\}\)\(4\)
3\. Style Consistency Score \(SCS\)SCS is a VLM\-based \(gemini\-3\-pro\-preview\) perceptual metric evaluating three dimensions on a 0\-10 scale\.

SCS=130​\(Svisual\+Slayout\+Saesthetic\)\\mathrm\{SCS\}=\\frac\{1\}\{30\}\\left\(S\_\{\\text\{visual\}\}\+S\_\{\\text\{layout\}\}\+S\_\{\\text\{aesthetic\}\}\\right\)\(5\)Where:

- •SvisualS\_\{\\text\{visual\}\}: Color palette, line styles, node shapes\.
- •SlayoutS\_\{\\text\{layout\}\}: Spatial arrangement, alignment, spacing\.
- •SaestheticS\_\{\\text\{aesthetic\}\}: Overall visual harmony and professional look\.

4\. CodeXQA AccuracyCodeXQA is the average accuracy across the three question types \(Counting, Identification, Relationship\)\.

CodeXQA=1N​∑i=1N𝕀​\(Match​\(Answeripred,Answerigt\)\)\\mathrm\{CodeXQA\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbb\{I\}\(\\mathrm\{Match\}\(\\text\{Answer\}\_\{i\}^\{\\text\{pred\}\},\\text\{Answer\}\_\{i\}^\{\\text\{gt\}\}\)\)Matching strategies include Exact Match, Inclusion, and Semantic Similarity\.

5\. SigLIP2 Score \(SigLIP2 Semantic Similarity Score\)This metric uses the SigLIP2 model to calculate semantic similarity between the original and generated images\.

SigLIP2=CosineSimilarity​\(SigLIP2​\(Original\),SigLIP2​\(Generated\)\)\\mathrm\{SigLIP2\}=\\mathrm\{CosineSimilarity\}\(\\mathrm\{SigLIP2\}\(\\text\{Original\}\),\\mathrm\{SigLIP2\}\(\\text\{Generated\}\)\)The model used isgoogle/siglip2\-so400m\-patch16\-512, with cosine similarity in the range 0–1\.

#### F\.2\.2Task 2: Editing Metrics

1\. XDRFR \(XML Decomposed Requirements Following Ratio\)XDRFR is the primary metric for instruction following; it calculates the pass rate of decomposed Yes/No questions\.

XDRFR=1M​∑i=1M𝕀​\(Answeri="Yes"\)\\mathrm\{XDRFR\}=\\frac\{1\}\{M\}\\sum\_\{i=1\}^\{M\}\\mathbb\{I\}\(\\text\{Answer\}\_\{i\}=\\text\{"Yes"\}\)WhereMMis the number of decomposed questions for the instruction\. The evaluation is performed purely on the XML text, avoiding visual rendering artifacts\.

2\. SCS for Task 2This metric is adapted to focus on style preservation during edits\.

SCSTask2=120​\(Sstyle\+Saesthetic\)\\mathrm\{SCS\}\_\{\\text\{Task2\}\}=\\frac\{1\}\{20\}\\left\(S\_\{\\text\{style\}\}\+S\_\{\\text\{aesthetic\}\}\\right\)Crucially, this metric does not penalize content changes \(which are intended\) but ensures thestyleremains consistent with the original diagram\.

3\.mxGraphXML Edit Distance \(XED\)XED calculates the edit distance between the originalmxGraphXML and the modifiedmxGraphXML, quantifying the magnitude of code\-level modifications\.

XED=LevenshteinDistance​\(OriginalmxGraphXML,ModifiedmxGraphXML\)\\mathrm\{XED\}=\\mathrm\{LevenshteinDistance\}\(\\text\{Original \{mxGraph\} XML\},\\text\{Modified \{mxGraph\} XML\}\)It uses the standard Levenshtein distance algorithm, based on character\-level comparison ofmxGraphXML strings\.

#### F\.2\.3Evaluation Metrics Formalization

We compute layout\-aligned measures between reference \(RR\) and candidate \(CC\) outputs\.

- •Style Consistency: attribute agreement rate across evaluated fields, Style=∑i𝟏​\[attri\(r\)=attri\(c\)\]\#​attributes\.\\mathrm\{Style\}=\\frac\{\\sum\_\{i\}\\mathbf\{1\}\\big\[\\text\{attr\}\_\{i\}^\{\(r\)\}=\\text\{attr\}\_\{i\}^\{\(c\)\}\\big\]\}\{\\\#\\text\{ attributes\}\}\.
- •Editability: proportion of outputs that \(i\) load without errors and \(ii\) support move/resize and re\-routing without breaking attachments\.
- •Instruction Compliance: ratio of directives satisfied, combining rule\-based checks with adjudication on ambiguous cases\.

A weighted aggregate prioritizes structural fidelity and completeness while incorporating style, editability, and compliance\.

### F\.3Validation and Reproducibility

We use a validator pipeline to ensure schema validity and deterministic metric computation\.

VCG\-Bench Validator Pipeline:

Algorithm: VCG\-Bench Validator Pipeline

1. 1\.Input: candidatemxGraphXMLCC, referencemxGraphXMLRR, directivesDD
2. 2\.Parse and validate schema\(C\)\(C\); report well\-formedness and loadability
3. 3\.RenderCCandRRunder identical settings for spatial comparison
4. 4\.Detect elements and align via category \+ proximity; perform bipartite matching
5. 5\.Compute IoU, completeness, style consistency
6. 6\.Run editability checks: move/resize nodes, re\-route connectors
7. 7\.Evaluate directive compliance fromDD\(presence/absence, counts, layout constraints\)
8. 8\.Output: per\-metric scores and weighted aggregate