Rethinking NL2VIS: A comparative survey of LLM-based natural-language-to-visualisation frameworks

Publication Type

Conference Proceeding Article

Publication Date

6-2026

Abstract

Large Language Models (LLMs) are increasingly used to generate charts directly from natural-language instructions, yet their performance remains inconsistent across chart types and visualisation complexities. This paper examines a set of framework models that report results on comparable chart-generation benchmarks, enabling a clearer view of how these systems perform relative to one another. Through this comparative analysis, we identify persistent weaknesses in current LLM-based charting—such as unreliable library behaviours, loose alignment between code and visual output, and brittle reasoning that frequently fails in the absence of corrective mechanisms. We also highlight the broad strategies adopted across frameworks to address these shortcomings. Taken together, our findings suggest that future progress in Natural-Language-to-Visualisation (NL2VIS) will depend less on scaling models or increasing architectural complexity and more on developing lean, well-orchestrated systems with a stronger intrinsic understanding of how language, data, and visual semantics interact.

Keywords

Framework Models, LLM, NL2VIS, Visualisation

Discipline

Artificial Intelligence and Robotics | Graphics and Human Computer Interfaces

Research Areas

Intelligent Systems and Optimization

Publication

Proceedings of the 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2026, Hong Kong, China, June 9-12

First Page

138

Last Page

155

Identifier

10.1007/978-981-92-2014-4_12

Publisher

Springer

City or Country

Cham

Additional URL

https://doi.org/10.1007/978-981-92-2014-4_12

This document is currently not available here.

Share

COinS