LLM-Based Visualization Evaluation: How Well Do Literacy-Stratified Personas Approximate Human Judgments?

Applications, bias, and safety2026arXivApproved editorial review

Authors: Swaroop Panda

Keywords: Visualization literacy, LLM-as-participant, Human simulation, Psychometric validation

Source: Open primary source (opens in a new tab)

1
Authors
7
Findings
10
Limitations
5
Evidence

Editorial summary

English

The paper introduces LSLE, a framework that asks a multimodal LLM to respond as a person with low, medium, or high visualization literacy. It tests Claude Sonnet 4.6 at temperature .7 with ten generations per item and condition. The study administers all 53 VLAT items and the five BeauVis items over three visualizations in both image and structured-text modes, and includes a no-persona condition. The main result is negative: the personas do not produce the expected human gradient. On image-mode VLAT, low/medium/high score .75/.73/.75 versus .69 without a persona and a single published human aggregate near .65. In text mode all conditions reach .97–.98 because the description supplies chart type, axes, values, legend, and visual features, removing much of the perceptual task. Agreement with human item-difficulty ordering is weak in image mode and strongest without a persona (Spearman .44 versus .31/.24/.23); text-mode correlations are negative or near zero. The high-minus-low gap does not track published human item discrimination (r=-.08 image; -.15 text). The paper's own pooled table shows that every persona is farther from the human reference than the no-persona model: JSD .328/.343/.338 versus .289. On BeauVis it preserves the coarse BeamTree < StarTree < Sunburst order but overestimates appeal and does not establish a literacy effect. The audit prevents treating this as validation against human strata. The study has no individual human responses and constructs no human low/medium/high groups: all three personas are compared with the same published mean, difficulty, and discrimination statistics. The VLAT discrimination index is also not an observed human stratum gap equivalent to synthetic high-minus-low performance. Prompts directly prescribe chart-specific capabilities, errors, and confidence, with no reported calibration procedure, prompt alternatives, or held-out test set. The VLAT thresholds are correctly copied from the original 54-item tryout (mean 34.72, SD 7.05), but that administration offered an `Omit` option and guessing-corrected scores, whereas LSLE uses raw accuracy on the final 53 items without documenting omission or answer-order handling. A further factual error concerns BeauVis: the paper states N=150, while the primary validation recruited 201 and analyzed 197 after four exclusions; its bibliography also changes first author Tingying He to Tong He. Although the method declares JSD and Earth Mover's Distance, it reports no EMD value. It also omits the full prompts despite referring to supplementary material, exact stimuli and text descriptions, raw outputs, human distributions used, code, parser, failures, complete API configuration, and the statistics needed to reproduce correlations, JSD, or t-tests. Ten completions per item produce coarse uncertain distributions without intervals or sensitivity analysis. The defensible contribution is therefore a cautionary negative benchmark: verbal literacy descriptions can distort a capable model's answers without recreating human differences. The recommendation to use LSLE only as a formative complement is appropriately cautious, but usefulness for screening new designs remains untested.

Español

El artículo presenta LSLE, un marco para pedir a un LLM multimodal que responda como una persona con alfabetización visual baja, media o alta. Lo prueba con Claude Sonnet 4.6, temperatura 0,7 y diez generaciones por ítem y condición. Evalúa los 53 ítems del VLAT y los cinco ítems BeauVis sobre tres visualizaciones, tanto a partir de la imagen como de una descripción textual estructurada; añade una condición sin persona. El resultado principal es negativo: las personas no producen el gradiente humano esperado. En VLAT con imagen, baja/media/alta obtienen 0,75/0,73/0,75 frente a 0,69 sin persona y un promedio humano agregado cercano a 0,65. Con texto todas quedan en 0,97–0,98 porque la descripción entrega tipo de gráfico, ejes, valores, leyenda y rasgos visuales, eliminando buena parte de la tarea perceptiva. La concordancia del orden de dificultad con humanos es débil en imagen y máxima sin persona (Spearman 0,44 frente a 0,31/0,24/0,23); con texto es negativa o casi nula. La diferencia alta-menos-baja no sigue la discriminación humana por ítem (r=-0,08 en imagen; -0,15 en texto). La propia tabla conjunta muestra que cada persona se aleja más de la referencia humana que el modelo sin persona: JSD 0,328/0,343/0,338 frente a 0,289. En BeauVis conserva el orden grueso BeamTree < StarTree < Sunburst, pero sobrevalora el atractivo y no demuestra un efecto de alfabetización. La auditoría impide interpretar esto como validación de estratos humanos. El estudio no dispone de respuestas individuales ni construye grupos humanos bajo/medio/alto: compara las tres personas contra el mismo promedio, dificultad y discriminación publicados. El índice de discriminación VLAT tampoco es una diferencia observada entre estratos equivalente a alta-menos-baja. Los prompts prescriben directamente capacidades, errores por tipo de gráfico y confianza, sin calibración descrita, alternativas de prompt ni conjunto de prueba separado. Los umbrales VLAT sí están bien copiados del ensayo original de 54 ítems (media 34,72; DE 7,05), pero aquel protocolo ofrecía `Omit` y corrección por azar, mientras LSLE usa exactitud bruta sobre los 53 ítems finales sin documentar omisiones ni orden de respuestas. Hay un error factual adicional: el paper atribuye N=150 a la validación BeauVis, pero el trabajo primario reclutó 201 y analizó 197 tras cuatro exclusiones; su bibliografía también cambia Tingying He por Tong He. Aunque declara JSD y Earth Mover's Distance, no presenta ningún valor EMD. Tampoco publica prompts completos pese a remitir a material suplementario, imágenes y descripciones exactas, respuestas, datos humanos empleados, código, parser, fallos, configuración completa de API ni estadísticos suficientes para reproducir correlaciones, JSD o t-tests. Con diez completaciones por ítem, las distribuciones son además discretas e inciertas, sin intervalos ni análisis de sensibilidad. La aportación defendible es por tanto una advertencia empírica: describir verbalmente niveles de alfabetización puede distorsionar las respuestas de un modelo competente sin recrear diferencias humanas. La recomendación de usar LSLE solo como complemento formativo es prudente, pero su utilidad para seleccionar diseños nuevos todavía no se evalúa.

Research question

Can a multimodal LLM conditioned with low, medium, and high visual literacy descriptions approximate the accuracy, difficulty order, discrimination, and judgment distributions observed in humans on VLAT and BeauVis better than the same model without persona?

Method

Three system prompts are constructed from thresholds of one standard deviation of the original VLAT trial. Claude Sonnet 4.6 responds to each item under low, medium, high, or no persona, with ten samples at temperature 0.7. Image/textual description and VLAT/BeauVis are crossed, and accuracy, difficulty correlation, JSD, a discrimination correlation, and paired t-tests against the baseline condition are reported. The audit reviews the 25 pages, the complete TeX, the BeauVis arithmetic, the VLAT and BeauVis primary sources, and the availability of artifacts.

Sample: A single proprietary model; three personas and one no-persona condition; two modes; two instruments; ten completions per item and condition. The published VLAT reference comes from 191 participants in the trial, but individual responses and human strata are not available. BeauVis analyzed 197 participants, not the 150 indicated by LSLE.

Findings

  • No low < medium < high order appears: on VLAT with image the personas score 0.75/0.73/0.75 and with text 0.98/0.97/0.97.
  • The no-persona condition has the highest correlation with human difficulty on image (rho=0.44) and lower global JSD than the three personas.
  • The synthetic high-minus-low difference does not reproduce human discrimination by item: r=-0.08 on image and r=-0.15 on text.
  • The textual mode creates a ceiling of 97-98% and does not preserve the human difficulty order, consistent with having eliminated much of the graphical decoding.
  • BeauVis preserves the coarse order of the three stimuli, but overestimates their human scores by approximately 0.6 to 1.4 points.
  • The twelve published BeauVis composites do reproduce the arithmetic mean of their five items, within rounding.
  • The direction of the published t-tests is adverse to LSLE: all personas significantly increase JSD relative to the model without persona.

Limitations

  • There are no corresponding low/medium/high human groups; the three personas are compared against a single historical aggregate.
  • The prompts directly encode the skills, failures, and confidence assessed, without reproducible calibration or out-of-sample testing.
  • The textual description provides values and features and is therefore not equivalent to reading the chart.
  • Only one model, one prompt per stratum, one temperature, and ten samples per item are studied.
  • No complete prompts, stimuli used, descriptions, responses, code, data, parser, errors, or executable configuration are published.
  • EMD is declared repeatedly but not reported; JSD, correlations, and p-values cannot be recalculated.
  • The t-test does not report pairing unit, n, degrees of freedom, statistic, intervals, or correction for three comparisons.
  • The BeauVis figure N=150 contradicts the primary source, which analyzes N=197.
  • The `Omit` option, chance correction, option randomization, and positional bias for VLAT are not documented.
  • Utility for screening new designs is recommended but is not subjected to an evaluation of subsequent human decisions or outcomes.

What the study does not establish

  • It does not demonstrate that LLM personas represent real human strata of visual literacy.
  • It does not demonstrate fidelity of graphical perception, metacognition, cognitive strategies, or individual variability.
  • It does not demonstrate that using persona improves human alignment; the published metrics show the opposite effect.
  • It does not demonstrate that preserving the order of three BeauVis stimuli depends on literacy or is equivalent to simulating human preferences.
  • It does not validate EMD, robust inferential significance, or independent reproducibility.
  • It does not yet validate LSLE as a tool for selecting designs in a real UI/UX or visualization workflow.

Traceability

Scope: Full text

Version: arXiv:2606.10095v1

Consulted source: https://arxiv.org/abs/2606.10095v1

Review: Codex twenty-five-page full-text visual, TeX, primary-human-baseline, psychometric, statistical, arithmetic and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Claude Sonnet 4.6 (identificador exacto y fecha de snapshot no informados)

Instruments and metrics

  • Visualization Literacy Assessment Test (VLAT, 53 ítems finales; umbrales tomados del ensayo de 54 ítems)
  • BeauVis, escala de placer estético de cinco ítems sobre BeamTree, StarTree y Sunburst

Data used

  • Estadísticos humanos publicados de la validación VLAT
  • Estadísticos humanos publicados de la validación BeauVis
  • Respuestas sintéticas no publicadas de Claude en modo imagen y descripción textual

Evidence and location

  • Metadata, version, and preprint condition: Official arXiv record 2606.10095v1, checked 2026-07-17
  • Method, tables, discussion, limitations, and absence of the promised supplement: arXiv v1, all twenty-five PDF pages and complete TeX source
  • Figures, Omit, chance correction, and difference between VLAT trial and validation: Lee, Kim and Kwon 2017 primary VLAT paper, pp. 556-558
  • Actual size of the BeauVis validation and original open materials: He et al. 2023 primary BeauVis paper, Section 7; 201 recruited, 197 analyzed
  • Audit of human strata, metrics, statistics, arithmetic, and reproducibility: reports/verification/article-303-lsle-human-strata-aggregate-baseline-persona-failure-statistical-and-artifact-audit.json