A CANVAS-NATIVE DIAGRAMMING FRAMEWORK

EvoDesign

Agentic Editable Diagram Creation
via Design Expertise Evolution

Tianfu Wang1Leilei Ding2Ziyang Tao2Yi Zhan2Zhiyuan Ma2Wei Wu2Yuxuan Lei2Junyang Wang2Yin Wu1Yizhao Xu3Hongyuan Zhu2Qi Liu2Nicholas Jing Yuan1Yanyong Zhang2Hui Xiong1,4
1 The Hong Kong University of Science and Technology (Guangzhou)2 University of Science and Technology of China3 Peking University4 The Hong Kong University of Science and Technology
Read the paper Code Coming soonExplore the method

Diagrams you can edit. Design expertise that evolves.

Read the abstract Cite this work
THE WORKFLOW, IN MOTION

Designed to stay editable.

Follow the process. Then make it yours.

From intent to editable canvasINTERACTIVE ILLUSTRATION
Illustrative EvoDesign workflow. Drag the five nodes, or focus a node and use arrow keys to move it.COORDINATED DESIGN AGENTSVerified experience → evolving expertiseCONTENT → STRUCTURE → STYLE → LAYOUT → REFINEMENT
Every element stays editableDrag a node to make it yours
Start with an idea.

Source content and your requirements set the intent for the diagram.

An interactive illustration of the framework. The research system will be released with the code.

01 / THE IDEA

Editable by design.
Better with experience.

EvoDesign is an agentic framework for editable diagram generation that coordinates structure, style, and layout agents and evolves reusable design expertise.

RepresentationObject-level editable canvas
FeedbackGeometric checks + VLM critique
LearningEvidence-guided expertise memory
Read the full paper

Abstract

High-fidelity diagram creation requires coordinated visual-spatial decisions over semantic topology, visual styling, and geometric layout. Existing methods face a representation gap: pixel-based generation offers limited object-level control, while code-based synthesis provides executable structure at the cost of intuitive manipulation. We introduce EvoDesign, an agentic framework for editable diagram creation through an intermediate canvas schema that is both machine-actionable and directly manipulable by users. EvoDesign first constructs a diagram manifest through coordinated structure, style, and layout agents, and then renders the manifest in a diagnostic verification environment. This environment produces objective defect signals and VLM-based critique, enabling localized refinement of the current diagram while also providing evidence for long-term design expertise evolution. The evolution mechanism distills refinement traces into a three-tier design expertise memory, where candidate strategies are linked to the verification evidence that supports them and are promoted into broader guidelines and principles only through repeated cross-context support. We further define CanvasBench, a canvas-recoverable benchmark and evaluation protocol with content, visual, and cognitive dimensions. Experiments across pixel, code-based, agentic, one-shot schema, and canvas-native methods show that EvoDesign provides the strongest overall balance across content fidelity, visual organization, and cognitive usability.

Abstract from the manuscript; project name updated to EvoDesign. Code release forthcoming.

02 / THE FRAMEWORK

Create. Verify. Evolve.

One editable representation connects
generation, refinement, and learning.

01

Structure

Identify entities, relations, groups, and hierarchy from the source content.

Semantic graph G
02

Style

Translate visual intent into consistent colors, shapes, typography, and emphasis.

Visual encoding S
03

Layout

Estimate element sizes, arrange nested groups, and route connections on the canvas.

Spatial configuration L
04

Verify & refine

Combine geometric checks with VLM critique to localize defects and repair affected objects.

Verified refinement ↺

Design expertise, accumulated.

Refinement traces become reusable knowledge,
with evidence preserved at every level.

01 / SPECIFICSample strategies
02 / CONTEXTUALDomain guidelines
03 / TRANSFERABLEGeneral principles
What makes expertise evolution evidence-guided?

Each candidate strategy is linked to the specific findings, edits, and outcomes that support it. Repeated support across compatible contexts promotes strategies into domain guidelines; support across diagram types promotes guidelines into general principles. Retrieval uses agent role, diagram type, and domain, while maintenance prunes, merges, promotes, or demotes memory items.

03 / A CLOSER LOOK

Repair the details.
Keep the intent.

When a diagram becomes crowded, EvoDesign updates the affected objects in the canvas manifest. Semantic structure can stay intact while spacing, routing, and visual grouping improve.

Localized editsObject-level controlDiagnostic feedback

Software development lifecycle with change control. Artifacts from the manuscript’s refinement case study (Figure 9); shown for qualitative inspection.

A Initial solution

Crowded subprocess & return paths

B After refinement

Clearer spacing & visual grouping

04 / THE BENCHMARK

Meet CanvasBench.

A benchmark built around what
an editable canvas can represent.

6,273Validated samples
21Diagram types
30Content domains
420Held-out test samples

Canvas-recoverable diagram–instruction pairs spanning real-world visual topologies. The 5,853-sample training/evolution split supports expertise accumulation; the balanced test split contains 20 samples per diagram type.

FlowchartsConcept mapsArchitectureMind mapsState diagramsTimelines+ 15 more
C

Content integrityFidelity · Concept logic · Relevance

V

Visual presentationAesthetics · Consistency · Flow · Space

G

Cognitive utilityEase · Accuracy · Self-contained explanation

05 / REPORTED RESULTS

Quality, from three perspectives.

The manuscript evaluates semantic fidelity, visual organization, and cognitive usability using VLM judges on a five-point scale.

Mean score within each dimension · GPT-5.5 backbone for non-pixel methods · higher is better
MethodRepresentationContent / 5Visual / 5Cognitive / 5
GPT-4o-ImagePixel-based2.3013.6673.080
LaTeX TikZCode-based3.4682.8563.199
DiagramAgentAgentic2.1322.7072.450
SciSketchAgentic3.8102.9523.307
Schema promptingOne-shot schema3.6872.4822.752
EvoDesignCanvas-native4.0993.6593.503

Dimension means are calculated from the 10 metrics in manuscript Table 1 (3 content, 4 visual, 3 cognitive). The pixel baseline uses GPT-4o-Image and stays fixed across backbone selections. These are manuscript-reported results, not an independent reproduction. See the paper for the evaluation protocol.

Showing GPT-5.5 · 6 methods · 3 dimensions

Download all reported metrics
+1.472

Space balanceover one-shot schema prompting

+1.106

Cognitive easeover one-shot schema prompting

GPT-5.5 backbone
Individual metrics on a 1–5 scale
Manuscript Table 1

06 / RESEARCH NOTES

A few useful
distinctions.

What the framework does,
and what the evidence supports.

How is EvoDesign different from image or code generation?

EvoDesign uses an intermediate canvas schema with semantic structure, visual style, and spatial layout. The result consists of individual objects that users can move, regroup, and restyle. Specialized agents operate on the same representation to create and locally repair a diagram.

Explore the framework
What does CanvasBench evaluate?

CanvasBench evaluates editable diagram generation using 6,273 validated samples across 21 diagram types and 30 content domains. Its 420-sample held-out test set covers content integrity, visual presentation, and cognitive utility through 10 metrics.

See the benchmark
How should I interpret the reported results?

The comparison reports dimension averages from manuscript Table 1, using VLM judges on a five-point scale. These are not independent reproductions or human preference scores. With GPT-5.5, EvoDesign scores higher on content and cognitive means than the displayed baselines; GPT-4o-Image has a slightly higher visual mean.

Inspect the comparison
Can I use the code or download CanvasBench today?

The research code is coming soon, and CanvasBench release details are forthcoming. This website’s draggable canvas illustrates the workflow; it does not run the research system. The paper is available on arXiv, and release updates will appear in the project repository.

Follow the repository

07 / CITATION

Build on
this work.

If EvoDesign or CanvasBench informs your research, please cite the paper.

arXiv:2604.09568

Tianfu Wang, Leilei Ding, Ziyang Tao, et al. 2026

EvoDesign: Agentic Editable Diagram Creation via Design Expertise Evolution

arXiv preprint · DOI: 10.48550/arXiv.2604.09568

BibTeX RIS

Import BibTeX or RIS into your reference manager.

View BibTeX 

@misc{wang2026evodesign,
  title = {{EvoDesign}: Agentic Editable Diagram Creation via Design Expertise Evolution},
  author = {Wang, Tianfu and
            Ding, Leilei and
            Tao, Ziyang and
            Zhan, Yi and
            Ma, Zhiyuan and
            Wu, Wei and
            Lei, Yuxuan and
            Wang, Junyang and
            Wu, Yin and
            Xu, Yizhao and
            Zhu, Hongyuan and
            Liu, Qi and
            Yuan, Nicholas Jing and
            Zhang, Yanyong and
            Xiong, Hui},
  year = {2026},
  eprint = {2604.09568},
  archivePrefix = {arXiv},
  primaryClass = {cs.HC},
  doi = {10.48550/arXiv.2604.09568},
  url = {https://arxiv.org/abs/2604.09568}
}

CONTINUE EXPLORING

Built to be built upon.

The EvoDesign code release is on the way.
Follow the repository for implementation updates.

Project repositoryCode coming soonResearch paperarXiv:2604.09568
CanvasBenchDataset release details forthcomingComing soon
REFINEMENT CASE STUDY

Initial solution

Software development lifecycle with change control · Manuscript Figure 9