[Paper] A2P-Vis: an Analyzer-to-Presenter Agentic Pipeline for Visual Insights Generation and Reporting

Source: arXiv - 2512.22101v1

Overview

The paper introduces A2P‑Vis, a two‑stage, multi‑agent system that turns raw tabular data into a polished, publication‑ready data‑visualization report. By pairing a Data Analyzer that automatically creates, evaluates, and scores visual insights with a Presenter that weaves those insights into a coherent narrative, the authors bridge the long‑standing gap between raw analytics and finished storytelling for data‑science practitioners.

Key Contributions

• End‑to‑end agentic pipeline that goes from dataset ingestion to a complete visual report without human “glue” work.
• Data Analyzer agent that:
  • Profiles the dataset and proposes a diverse set of visualization directions.
  • Generates executable plotting code (e.g., Matplotlib, Altair) and runs it automatically.
  • Uses a legibility checker to filter out low‑quality figures (e.g., unreadable axes, overlapping labels).
  • Extracts candidate insights from each chart and scores them on depth, correctness, specificity, and actionability.
• Presenter agent that:
  • Orders the top‑ranked insights into logical sections.
  • Crafts chart‑grounded narrative paragraphs, adds justified transitions, and iteratively revises for clarity and consistency.
• Insight scoring framework that quantifies “insight quality” along multiple dimensions, enabling the system to prioritize the most valuable findings.
• Open‑source demo and dataset (https://www.visagent.org/api/output/f2a3486d-2c3b-4825-98d4-5af25a819f56) for reproducibility and community extension.

Methodology

1. Data Ingestion & Profiling – The Analyzer reads a CSV/Excel file, computes basic statistics, and detects column types (categorical, numeric, datetime).
2. Visualization Generation – Using a prompt‑driven LLM, the Analyzer suggests several chart types (bar, scatter, heatmap, etc.) and emits corresponding Python code. The code is executed in a sandbox; generated figures are stored.
3. Legibility Checking – A lightweight vision model (or rule‑based heuristics) evaluates each figure for readability (e.g., sufficient contrast, non‑overlapping labels). Poor figures are discarded.
4. Insight Extraction & Scoring – For each retained chart, the Analyzer prompts an LLM to describe the visual pattern and then applies a scoring rubric that measures:
   • Depth: how far the insight goes beyond surface statistics.
   • Correctness: alignment with the underlying data.
   • Specificity: avoidance of vague statements.
   • Actionability: whether the insight suggests a concrete next step.
5. Narrative Construction – The Presenter receives the top‑ranked insights and their associated charts. It orders them (e.g., exploratory → explanatory), writes paragraph‑level text anchored to each chart, adds transitional sentences, and runs a second LLM pass for polishing.
6. Iterative Revision – The system loops back to fix inconsistencies (e.g., mismatched terminology) and to ensure a uniform style throughout the report.

Results & Findings

• Quality boost: Compared against a baseline single‑agent system that only generates charts, A2P‑Vis achieved a +27 % increase in human‑rated report quality (clarity, insightfulness, and visual appeal).
• Insight depth: The scoring rubric filtered out 38 % of generated insights as “shallow,” leaving a concise set of high‑impact findings.
• Diversity of visualizations: The Analyzer produced an average of 4.3 distinct chart types per dataset, covering both univariate and multivariate relationships.
• Human evaluation: In a user study with 15 data analysts, 80 % preferred the A2P‑Vis reports over manually assembled notebooks, citing faster comprehension and better storytelling.
• Runtime: End‑to‑end generation for a 10 k‑row dataset completed in under 2 minutes on a single GPU‑enabled workstation.

Practical Implications

• Rapid prototyping – Data engineers can feed raw logs or business metrics into A2P‑Vis and receive a ready‑to‑share report, cutting exploratory analysis time from hours to minutes.
• Automated reporting for dashboards – The pipeline can be scheduled to produce periodic visual briefs (e.g., weekly sales performance) without manual chart selection or narrative writing.
• Education & onboarding – New analysts can study the generated reports to learn best‑practice visual storytelling and insight formulation.
• Integration with CI/CD – Teams building data products can embed A2P‑Vis as a post‑processing step, automatically documenting model performance or data drift.
• Customization hooks – Because the Analyzer and Presenter are modular, organizations can swap in domain‑specific scoring functions or corporate style guides, ensuring the output aligns with internal standards.

Limitations & Future Work

• Reliance on LLM correctness – The system inherits the hallucination risk of large language models; occasional mis‑interpreted patterns still slip through the scoring filter.
• Domain‑specific nuance – The current scoring rubric is generic; specialized fields (e.g., genomics, finance) may need tailored metrics for “actionability.”
• Scalability to massive datasets – While the demo handles up to ~100 k rows comfortably, larger data would require sampling strategies or distributed execution.
• User control – Presently the pipeline operates autonomously; future versions could expose knobs for analysts to steer visualization direction or narrative tone.
• Evaluation breadth – The paper’s user study is limited in size and diversity; broader industrial trials would better validate real‑world impact.

Bottom line: A2P‑Vis demonstrates that coupling a quality‑assured visual analyzer with a narrative‑focused presenter can deliver end‑to‑end, AI‑driven data‑storytelling that’s both technically sound and ready for business consumption. As the ecosystem of LLM‑augmented tools matures, pipelines like this could become the default “report‑as‑code” paradigm for data teams.

Authors

• Shuyu Gan
• Renxiang Wang
• James Mooney
• Dongyeop Kang

Paper Information

• arXiv ID: 2512.22101v1
• Categories: cs.LG, cs.AI, cs.CL
• Published: December 26, 2025
• PDF: Download PDF