CoDA: Agentic Systems for Collaborative Data Visualization Zichen Chen, Jiefeng Chen, Sercan O. Arik, Misha Sra, Tomas Pfister, Jinsung Yoon International Conference on Learning Representations (ICLR 2026)
CoDA (Collaborative Data-visualization Agents) is a multi-agent framework that turns natural language queries into publication-quality visualizations. Instead of treating visualization as a monolithic code-generation problem, CoDA distributes the work across 8 specialized LLM agents in four sequential phases — Understanding, Planning, Generation, and Self-Reflection — that iteratively refine the output until explicit quality thresholds are met.
| Benchmark | CoDA | Improvement over Prior SOTA |
|---|---|---|
| MatplotBench | 79.5% | +24.5% |
| Qwen Code Interpreter | 89.0% | new state-of-the-art |
| DA-Code | 39.0% | +2× prior best |
| Human Evaluation (Elo) | 1701 | outperforms all baselines |
CoDA employs 8 specialized LLM agents across 4 phases. A global TODO list is generated by QueryAnalyzer at the start and is propagated through every subsequent agent; VisualEvaluator uses image-level analysis to verify completion before the workflow terminates.
| Phase | Agent | Responsibility |
|---|---|---|
| 1 — Understanding | QueryAnalyzer | Parses intent, expands ambiguous queries, produces a structured TODO checklist |
DataProcessor | Loads heterogeneous data files, infers schema and statistics | |
| 2 — Planning | VisualizationMappingAgent | Assigns data columns to visual roles (axes, color, size) and selects chart type |
DesignExplorer | Specifies color scheme, layout, typography, and accessibility choices | |
| 3 — Generation | SearchAgent | Retrieves real matplotlib gallery examples to bias the code generator |
CodeGenerator | Produces clean, runnable matplotlib code grounded in the design spec | |
| 4 — Self-Reflection | DebugAgent | Executes the code, diagnoses runtime errors, and applies targeted fixes |
VisualEvaluator | Scores readability, aesthetics, and UX; triggers refinement if below threshold |
CoDA consistently produces complete, accurate visualizations scoring 90–95/100, while baselines frequently produce broken outputs scoring 0–45/100.
| Task | Benchmark | Score | CoDA Output | Ground Truth |
|---|---|---|---|---|
| NBA team performance trends — multi-line chart | DA-Code | 100 / 100 |  |  |
| Steam game ratings — scatter plot with trend line | DA-Code | 100 / 100 |  |  |
| Hierarchical data — sunburst chart | MatplotBench | 100 / 100 |  |  |
CoDA's self-reflection loop detected layout and labeling errors across 3 iterations, then automatically recovered at iteration 4 — demonstrating the robustness of quality-driven refinement.
| Iter 3 (Failed) | Iter 4 (Recovered) |
|---|---|
 |  |
Requirements: Python ≥ 3.8 and API credentials for at least one LLM provider.
git clone https://github.com/google-research/agentic-visualization.git
cd agentic-visualization
pip install -r requirements.txtConfigure your API keys:
cp .env.example .env
# Fill in OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, etc.importcodaresult=coda.plot(
query="Create a sunburst chart showing browser market share by version.",
data="path/to/data.csv", # CSV, JSON, Excel, SQLite, or a directoryoutput="result.png",
)
print(f"Quality: {result.quality_score:.2f} — saved to {result.output_file}")fromcodaimportWorkflowfromcoda.workflow.orchestratorimportWorkflowConfigurationcfg=WorkflowConfiguration(
quality_threshold=0.9,
max_iterations=3,
enable_search_agent=True,
output_directory="outputs/",
)
wf=Workflow(config=cfg, query_id="q001")
result=wf.execute_workflow(
query="Plot monthly active users with a 90-day rolling average.",
data_input="data/mau.csv",
workflow_context={"original_data_path": "data/mau.csv"},
)
print(f"Success: {result.final_success} | Quality: {result.final_quality_score:.2f}")
# Output image: outputs/<model>_query_q001/query_1/final_result.pngbash prepare_benchmark.sh
# Downloads MatPlotAgent benchmark into benchmark_data/python run_benchmark.py \
--start 1 --end 100 \
--model-name openai/gpt-4o \
--search-model-name openai/gpt-4o-miniResults are saved to benchmark_outputs/<model>_query_<id>/.
python run_evaluation.py \
--results-dir benchmark_outputs \
--start 1 --end 100 \
--model-name openai/gpt-4o \
--vision-model-name openai/gpt-4o \
--processes 10Model tags follow the LiteLLM naming convention (provider/model-name). Any provider supported by LiteLLM — OpenAI, Anthropic, Google Gemini, Groq, etc. — works out of the box.
@inproceedings{chen2026coda,
title = {{CoDA}: Agentic Systems for Collaborative Data Visualization},
author = {Chen, Zichen and Chen, Jiefeng and Arik, Sercan O. and Sra, Misha and Pfister, Tomas and Yoon, Jinsung},
booktitle = {International Conference on Learning Representations},
year = {2026},
url = {https://coda-agent.github.io/CoDA/}
}Apache 2.0 License — see LICENSE for details.
This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.
This project is intended for demonstration purposes only. It is not intended for use in a production environment.