When Do LLM Personas Support Visualization Design? A Cross-Model Study of Color Assignment and Chart Choice

Jul 16, 2026

Large language model personas are increasingly used to approximate diverse users during early-stage visualization design, yet it remains unclear whether persona-conditioned outputs reflect stable personality effects or artifacts from model choice and task framing. This study examines color assignment for abstract versus concrete concepts and chart-idiom preference ratings across task contexts. Across GPT-4o-mini, GPT-4.1-mini, and GPT-5-mini, personality-color coupling varies by model and concept type, while chart choice is more context-driven than personality-driven.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

Abstract

Introduction

Background and Related Work

© This is the author’s version of the article When Do LLM Personas Support Visualization Design? A Cross-Model Study of Color Assignment and Chart Choice Shahreen Salim* Stony Brook University Klaus Mueller† Stony Brook University arXiv :2607 .02455v 1 [ cs .HC] 2 Jul 2026 ABSTRACT Large language model personas are increasingly used to approximate diverse users during early-stage visualization design, but it remains unclear whether persona-conditioned outputs reflect stable personality effects or artifacts of model choice and task framing. We examine this question across two visualization-relevant tasks: color assignment for abstract and concrete concepts, and chartidiom preference ratings across task contexts. Using 43 Big Five profiles across GPT-4o-mini, GPT-4.1-mini, and GPT-5-mini, we find that personality-color coupling is highly model-configuration dependent: absent in GPT-4o-mini for all six concepts, consistent in GPT-4.1-mini across all six, and partial in GPT-5-mini for two of six. Concept type further shapes the signal: for abstract concepts, personality explains more hue variance than model identity, while concrete concepts show smaller and comparable effects. In chart choice, trait-aligned cluster aggregation produces stable topidiom rankings across all nine cluster-context combinations, but ano-persona baseline recovers the same top choice in 8 of 9 modelcontext cells, indicating that task context drives rank-1 selection more than personality. These findings position LLM personas as exploratory probes for visualization design, not substitutes for human participants, and motivate multi-model testing, concept-typedisaggregation, and no-persona baselines in future studies. Index Terms: Large language models, Color design, Visual analytics, Personalization, Human computer interaction 1 INTRODUCTION LLM personas are increasingly used to simulate diverse user profiles during early-stage visualization research [9, 20] . The appeal is clear: if conditioning a model on a Big Five profile [10] produces systematic differences in visualization choices, researchers could use personas to explore design alternatives before running costly human studies. This use case is especially attractive for personalization problems, where color choices, chart preferences, and interpretation strategies may vary across users. However, the same convenience creates a methodological risk. Persona-conditioned outputs may reflect model-specific artifacts, prompt defaults, or task semantics rather than stable personality effects. We examine this risk in two visualization-relevant settings. Experiment 1 tests whether Big Five persona conditioning changes the colors that LLMs assign to abstract and concrete concepts. Experiment 2 tests whether the same persona profiles produce stable chart-idiom preferences across task contexts, extending the chartpreference setting of Alves et al. [2] . Across both experiments, we compare GPT-4o-mini, GPT-4.1-mini, and GPT-5-mini using the same 43 unique Big Five profiles; Experiment 2 resamples these profiles to 60 persona-conditioned runs to approximate Alves et al.’s sample size. * e-mail: [ssalimaunti@cs.stonybrook.edu](ssalimaunti@cs.stonybrook.edu)[ ](ssalimaunti@cs.stonybrook.edu)†[e-mail: mueller@cs.stonybrook.edu](e-mail: mueller@cs.stonybrook.edu) Our results show that persona effects are not a single property of the prompt. In color assignment, personality-color coupling depends strongly on model configuration and concept type: the signal is absent in GPT-4o-mini, consistent in GPT-4.1-mini, and partial in GPT-5-mini, with stronger persona-driven variation for abstract concepts than concrete ones. In chart choice, trait-aligned cluster aggregation stabilizes rankings, but a no-persona baseline shows that rank-1 idiom selection is mostly context-driven. These findings support a methodological argument for persona-based visualization studies: test multiple model configurations, sepa

FAQs

Do LLM personas consistently produce stable personality-driven visualization effects across models?

No. Personality-color coupling depends strongly on model configuration: absent in GPT-4o-mini for all concepts, consistent in GPT-4.1-mini across all concepts, and partial in GPT-5-mini for some concepts.

How does concept type affect persona-driven color assignment?

Personality explains more hue variance for abstract concepts than for concrete ones. Concrete concepts show smaller and more comparable effects between personality and model identity.

Is chart idiom preference mainly driven by personality or by task context?

Task context dominates rank-1 selection. While trait-aligned cluster aggregation yields stable top idiom rankings across cluster-context combinations, a no-persona baseline recovers the same top choice in 8 of 9 model-context cells.

Jul 16, 2026

Upload by Emma Wilson