ATWL: A Formal Language for Representing, Comparing, and Reusing Visual Analytics Workflows
- Fraunhofer Institute IAIS & Lamarr Institute for Machine Learning and Artificial Intelligence, Sankt Augustin, Germany
- City St George's, University of London, UK
- University of Zurich, Switzerland
- University of Stuttgart, Germany
Language Definition
-
ATWL Language Definition
Complete specification: artifact types, transform intents, control structures, and syntax reference.
Extraction Instructions
Guidelines for extracting workflows from research papers into ATWL.
Reviewing Instructions
Guidelines for reviewing ATWL representations for correctness.
Workflow Library
Library (HTML, 54 workflows)
Individual pages for each workflow with summaries and full ATWL representations, currently 54 workflows.
Library PDF compilation (Initial 18 workflows)
An initial subset (18 workflow representations) in a single document.
ATWL Workflow Lens
54 visual analytics workflows formalised in ATWL, the 25 binary features that characterise them, and how the workflows compare. State of 2 October 2026.
All features are computed deterministically from the ATWL text.
Library explorer
Visual analytics explorer of visual analytics workflows.
Graph measures, encodings, and diagrams are computed live from each page's own ATWL source.
ATWL support infrastructure
-
Pre-Trained LLM Agents: Agents & Usage Instructions
Session logs (JSON) for initializing LLM agents with ATWL expertise: Extractor, Reviewer, Diagrammer, and Recommender.
-
ATWL Syntax Checker
A page for syntax checking, correcting, and visualizing ATWL specifications.
Example of ATWL code with errors for testing purposes: cluster-calendar_corrupted.atwl
Workflow Recommendation Experiment, 2 tasks × 2 models × 7 prompting conditions (R1)
-
Full-scale recommendation experiment for the revised ATWL paper
This is the full-scale follow-up (R1) to the 2×2 pilot in experiments, run for the revised version of the paper. Where the pilot compared Formal vs. Prose library representations on two problems with one model, this study crosses 2 design tasks × 2 LLMs (Claude Sonnet 5, GPT-5.6-Terra) × 7 prompting conditions (plus two exploratory extensions), and compares the resulting workflow designs quantitatively and qualitatively: the ATWL designs with the 25 formal features of the paper and the ATWL syntax checker, the prose designs by coding under a written protocol.
Re-Extraction Check (R1)
-
How much does a workflow representation change when the workflow is extracted again?
Five library workflows (UTOPIAN, DPVis, EventThread, NodeTrix, SOMFlow) were extracted again with Claude Sonnet 5 and GPT-5.6-Terra. Each version was reviewed by the other model, revised once, and compared with the library version and with the other re-extraction. The backbone of a workflow (closed cycles, interpretation, judgement, and decision) is reproduced; details and the feature profiles of single workflows vary, so representations should be checked by the authors of the workflows. The page gives the procedure, the results, all ATWL representations, the review reports, and the data and scripts of the analysis.
Further materials from the initial submission
2×2 Recommendation Experiment (Old, from the initial submission)
-
Experiment Sessions & Materials
2×2 factorial experiment comparing Prose vs. Formal library representations on two problems (bike-sharing, VIS topic evolution). Includes session logs (JSON + HTML), ATWL workflow specifications, a narrated slide show, and two Jupyter notebooks. See Appendix E of the paper.
Reports & Results (Old, from the initial submission)
-
LLM Experiment Reports & Recommendations
Reports on LLM-based extraction, review, and recommendation experiments; systematic review of 16 workflows; and two detailed workflow recommendations (bike-sharing, diploma project).
Other Resources (Old, from the initial submission)
-
ATWL-to-HTML Converter Session
LLM session log for converting ATWL representations into styled HTML pages (used to generate the library pages).