Research map / Multi-agent systems and games / LLM agents
CoDA: Agentic Systems for Collaborative Data Visualization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoDA, a multi-agent framework (query analyzer, data processor, VizMapping, search, design explorer, code generator, debugger, visual evaluator) that generates and iteratively refines matplotlib visualizations from natural language and metadata, outperforming MatplotAgent, VisPath and CoML4VIS on MatplotBench and Qwen Code Interpreter using gemini-2.5-pro.
Problem
Automating data visualization from natural language over complex, multi-file datasets with code errors and iterative refinement needs; existing single- and multi-agent systems oversimplify by focusing on initial query parsing.
Approach
A self-evolving multi-agent pipeline that analyzes metadata schemas and statistics instead of raw data, maintains a global TODO list, retrieves code examples via a search agent, and uses image-based visual evaluation to route quality issues back to upstream agents.
Key results
- On MatplotBench, CoDA reaches 79.5% OS vs. 55.0% for MatplotAgent, 38.0% for VisPath and 53.0% for CoML4VIS (gemini-2.5-pro).
- On Qwen Code Interpreter visualization, CoDA reaches 89.0% OS vs. 81.6% for the best baseline VisPath.
- On DA-Code, CoDA reaches 39.0% overall score vs. 19.23% for DA-Agent with gemini-2.5-pro.
Surprising findings
- Removing the search agent lowers EPR by 9.0%, showing retrieval of plotting examples mainly helps executability.
- Gemini-2.5-flash backbone achieves 77.7% OS, only 1.8% below gemini-2.5-pro.
Limitations
- High computational overhead from multi-turn agent communication (about 14.8 LLM calls and 50k tokens per query).
- Evaluation relies on an LLM judge (gemini-2.5-pro) and comparison to ground-truth plots; the paper notes no human evaluation.
- Single-model backbone experiments are limited to proprietary Gemini and Claude models; no open-weight evaluation.
Experimental setup
Datasets: MatplotBench Qwen Code Interpreter benchmark (visualization subset) DA-Code
Benchmarks: MatplotBench Qwen Code Interpreter (visualization) DA-Code (vis)
Metrics: Execution Pass Rate Visualization Success Rate Overall Score token count LLM calls
Baselines: MatplotAgent VisPath CoML4VIS DA-Agent
Builds on: MatplotAgent VisPath CoML4VIS DA-Agent
Closest papers by meaning
- VisInteract: Towards Dynamic Interactive Text-to-Visualization under Imperfect Queries (NeurIPS 2026 · similarity 0.86)
- AutoData: A Multi-Agent System for Open Web Data Collection (NeurIPS 2025 · similarity 0.86)
- Multimodal DeepResearcher: Generating Text-Chart Interleaved Reports from Scratch with Agentic Framework (AAAI 2026 · similarity 0.85)
- Spider2-V: How Far Are Multimodal Agents From Automating Data Science and Engineering Workflows? (NeurIPS 2024 · similarity 0.85)
- Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling (CVPR 2026 · similarity 0.85)
- DV-World: Benchmarking Data Visualization Agents in Real-World Scenarios (ICML 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.