AgentViz: An Autonomous Multi-Agent Collaborative Framework for Hallucination-Free Natural Language Data Visualization on the nvBench Benchmark

Systems for Natural Language to Visualization (NL2VIS) allow non-expert users to automatically perform exploratory visual analytics over complex multi-table databases. However, current Large Language Model (LLM) single-turn architectures have limitations including visual hallucinations, suboptimal perceptual channel selection, and failures on ambiguous or multi-table join queries. In this paper, we propose AgentViz, an autonomous multi-agent collaborative framework for hallucination-reduced, perceptually optimized data visualization. The architecture decomposes the visualization pipeline into four cooperating agents: (1) a Data Profiler Agent that extracts statistical moments, semantic column data types, and correlation topologies; (2) a Semantic Disambiguation and Hypothesis Agent that identifies analytical hypotheses and resolves semantic constraints; (3) a Perceptual Visual Design Agent that leverages Cleveland-McGill graphical perception rankings to generate declarative Vega-Lite specifications; and (4) a Quality Critic and Verification Agent that performs closed-loop verification to identify field hallucinations and visual distortions. Evaluated on 25,750 natural language queries across 105 domains in the nvBench 2.0 benchmark, AgentViz reports an overall Visualization Exact Match Accuracy of 93.65%, a 10-fold reduction in visual hallucinations (1.15% versus 11.80%), and 89.20% accuracy on ambiguous multi-table queries, with an inference latency of 1.14 seconds per visualization. The framework aims to provide a verifiable approach to automated visual analytics and AI-assisted data visualization.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24360894
Primary Topic
Data Visualization and Analytics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

AgentViz: An Autonomous Multi-Agent Collaborative Framework for Hallucination-Free Natural Language Data Visualization on the nvBench Benchmark

Kanishka M
Zenodo (CERN European Organization for Nuclear Research)
Data Visualization and Analytics
preprint

AgentViz: An Autonomous Multi-Agent Collaborative Framework for Hallucination-Free Natural Language Data Visualization on the nvBench Benchmark

Kanishka M
preprint en

Abstract

Systems for Natural Language to Visualization (NL2VIS) allow non-expert users to automatically perform exploratory visual analytics over complex multi-table databases. However, current Large Language Model (LLM) single-turn architectures have limitations including visual hallucinations, suboptimal perceptual channel selection, and failures on ambiguous or multi-table join queries. In this paper, we propose AgentViz, an autonomous multi-agent collaborative framework for hallucination-reduced, perceptually optimized data visualization. The architecture decomposes the visualization pipeline into four cooperating agents: (1) a Data Profiler Agent that extracts statistical moments, semantic column data types, and correlation topologies; (2) a Semantic Disambiguation and Hypothesis Agent that identifies analytical hypotheses and resolves semantic constraints; (3) a Perceptual Visual Design Agent that leverages Cleveland-McGill graphical perception rankings to generate declarative Vega-Lite specifications; and (4) a Quality Critic and Verification Agent that performs closed-loop verification to identify field hallucinations and visual distortions. Evaluated on 25,750 natural language queries across 105 domains in the nvBench 2.0 benchmark, AgentViz reports an overall Visualization Exact Match Accuracy of 93.65%, a 10-fold reduction in visual hallucinations (1.15% versus 11.80%), and 89.20% accuracy on ambiguous multi-table queries, with an inference latency of 1.14 seconds per visualization. The framework aims to provide a verifiable approach to automated visual analytics and AI-assisted data visualization.

Zenodo (CERN European Organization for Nuclear Research)
Vellore Institute of Technology University (IN)
Data Visualization and Analytics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon