OpenNovelty.org

CoDA: Agentic Systems for Collaborative Data Visualization

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.3Download Report PDF
LLMmulti-agent systemvisualization
Abstract:

Automating data visualization from natural language is crucial for data science, yet current systems struggle with complex datasets containing multiple files and iterative refinement. Existing approaches, including simple single- or multi-agent systems, often oversimplify the task, focusing on initial query parsing while failing to robustly manage data complexity, code errors, or final visualization quality. In this paper, we reframe this challenge as a collaborative multi-agent problem. We introduce CoDA, a multi-agent system that employs specialized LLM agents for metadata analysis, task planning, code generation, and iterative reflection. We formalize this pipeline, demonstrating how metadata-focused analysis bypasses token limits and quality-driven refinement ensures robustness. Extensive evaluations show CoDA achieves substantial gains in the overall score, outperforming competitive baselines by up to 41.5%. This work demonstrates that the future of visualization automation lies not in isolated code generation but in integrated, collaborative agentic workflows.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces CoDA, a multi-agent system employing specialized LLM agents for metadata analysis, task planning, code generation, and iterative reflection to automate data visualization from natural language. It resides in the 'Multi-Agent LLM Systems' leaf of the taxonomy, which contains only four papers total, including CoDA itself. This represents a relatively sparse but rapidly growing research direction within the broader field of natural language-driven visualization, suggesting the multi-agent paradigm is still emerging compared to more established single-agent or direct prompting approaches.

The taxonomy reveals that CoDA's immediate neighbors include other multi-agent frameworks like nvAgent and VizGen, which similarly decompose visualization tasks into modular sub-tasks handled by specialized agents. The parent branch, 'Large Language Model-Based Translation Approaches', also encompasses 'Direct LLM Prompting and Code Generation' (five papers) and 'LLM-Based Grammar-Agnostic Generation' (two papers), indicating alternative architectural strategies. Adjacent branches such as 'Conversational Visualization Systems' and 'Exploratory Data Analysis Interfaces' address iterative refinement and dialogue, but through different mechanisms than multi-agent orchestration, highlighting CoDA's focus on agent collaboration rather than conversational interaction patterns.

Among the 30 candidates examined, the contribution-level analysis shows mixed novelty signals. The core 'CoDA multi-agent system' contribution examined 10 candidates and found 3 potentially refuting prior works, suggesting substantial overlap with existing multi-agent architectures. Similarly, 'reframing visualization as collaborative multi-agent problem' examined 10 candidates with 2 refutable matches, indicating this framing has precedent in the limited search scope. However, the 'formalization of metadata-centric preprocessing and iterative reflection pipeline' examined 10 candidates with zero refutable matches, suggesting this specific technical contribution may be more distinctive within the examined literature.

Given the limited search scope of 30 semantically similar candidates, this assessment captures overlap within the most directly related prior work but cannot claim exhaustive coverage of the broader visualization automation literature. The multi-agent leaf's sparsity (four papers) and the relatively high refutation rates for two contributions suggest CoDA builds incrementally on established multi-agent patterns, while its metadata-centric formalization may offer more novel technical detail. A more comprehensive search would be needed to assess whether similar metadata preprocessing or reflection mechanisms exist in adjacent research areas.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
5
Refutable Paper

Research Landscape Overview

Core task: automating data visualization from natural language queries. The field has evolved into a rich ecosystem with several major branches. Natural Language to Visualization Translation Systems form the technical core, encompassing rule-based parsers, neural sequence-to-sequence models, and increasingly sophisticated large language model-based approaches that can generate visualization code or specifications directly from user queries. Domain-Specific and Modality-Specific Visualization Systems tailor these techniques to particular data types (trajectories, medical records, sign language) or specialized domains, while Interactive and Conversational Visualization Interfaces enable iterative refinement through dialogue. Supporting branches include Natural Language Generation for Visualizations (producing textual descriptions of charts), infrastructure work on benchmarks and datasets, and a growing body of Applied Systems targeting real-world deployments in healthcare, business intelligence, and other sectors. Survey and Review Literature helps synthesize progress across these diverse threads. Recent momentum has concentrated in multi-agent LLM systems, where multiple specialized agents collaborate to handle query understanding, data processing, and chart generation. CoDA[0] exemplifies this trend by orchestrating agents for complex visualization tasks, sitting alongside other multi-agent frameworks like nvAgent[15] and VizGen[22] that similarly decompose the problem into modular sub-tasks. This contrasts with earlier single-model approaches such as Chat2vis[1] or LIDA[23], which relied on monolithic LLM prompting. The multi-agent paradigm offers improved modularity and error recovery but introduces coordination overhead. Meanwhile, works like Speech to Visualization[3] extend input modalities beyond text, and evaluation frameworks such as VisEval[5] provide systematic benchmarks to compare these varied architectures. CoDA[0] distinguishes itself within this cluster by emphasizing agent collaboration patterns and task decomposition strategies that balance flexibility with interpretability, positioning it as a representative example of how the field is leveraging LLM capabilities through structured orchestration rather than end-to-end generation alone.

Claimed Contributions

CoDA multi-agent system for data visualization

The authors introduce CoDA, a collaborative multi-agent system that employs specialized LLM agents to automate data visualization from natural language queries. The framework decomposes the task into understanding, planning, generation, and self-reflection phases, using metadata-focused analysis to bypass token limits and quality-driven refinement to ensure robustness in handling complex, multi-file datasets.

10 retrieved papers
Can Refute
Reframing visualization as collaborative multi-agent problem

The authors reconceptualize automated data visualization as a collaborative problem-solving endeavor rather than a monolithic single-pass process. This paradigm shift employs specialized LLM agents with distinct professional personas that use structured communication and quality-driven feedback loops to decompose queries, process data, and iteratively refine outputs.

10 retrieved papers
Can Refute
Formalization of metadata-centric preprocessing and iterative reflection pipeline

The authors formalize a pipeline that uses metadata-centric preprocessing to summarize data structures without full data loading, thereby circumventing LLM context window limits. The framework incorporates iterative reflection through image-based evaluation to verify completion from a human perspective, ensuring visualization quality through feedback loops.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

CoDA multi-agent system for data visualization

The authors introduce CoDA, a collaborative multi-agent system that employs specialized LLM agents to automate data visualization from natural language queries. The framework decomposes the task into understanding, planning, generation, and self-reflection phases, using metadata-focused analysis to bypass token limits and quality-driven refinement to ensure robustness in handling complex, multi-file datasets.

Contribution

Reframing visualization as collaborative multi-agent problem

The authors reconceptualize automated data visualization as a collaborative problem-solving endeavor rather than a monolithic single-pass process. This paradigm shift employs specialized LLM agents with distinct professional personas that use structured communication and quality-driven feedback loops to decompose queries, process data, and iteratively refine outputs.

Contribution

Formalization of metadata-centric preprocessing and iterative reflection pipeline

The authors formalize a pipeline that uses metadata-centric preprocessing to summarize data structures without full data loading, thereby circumventing LLM context window limits. The framework incorporates iterative reflection through image-based evaluation to verify completion from a human perspective, ensuring visualization quality through feedback loops.

Table of Contents