This preprint asks whether persona labels alter Qwen3-VL:8B captions, perception tags, and sentiment justifications for 50 PerceptSent urban images. Binary gender, income, and political labels plus three author-chosen personality archetypes form 24 profiles, each repeated 50 times. All persona calls use the same model, temperature 0.1, seed 42, and think=True; two no-persona variants provide one output per image. The paper analyzes 59,708 persona rows and 100 no-persona rows with all-MiniLM-L6-v2 cosine similarity, tag-set Jaccard similarity, per-image within/cross comparisons, profile-pair Pearson correlation, and BERTopic. Captions are highly similar across profiles. Justifications show reported within-group advantages for income of 0.062 and politics of 0.044, both p<0.001; personality is 0.023, p=0.052, and gender 0.002, p=0.702. No caption or tag dimension reaches p<0.05. Justification and tag divergence correlate at r=0.67. The defensible finding is narrow: with this model and prompt, demographic instructions leave more signal in the field explicitly required to speak in the persona's voice than in the field explicitly required to remain objective and persona-independent. The full artifact audit exposes major omitted boundaries. The factorial design requires 60,000 persona rows, but 292 failed. Failures are item-clustered: 185 occur on one image, 48 on another, and 21 on a third. Failure rates also vary by profile label, yet the paper says every agent annotated all 50 images and provides no missing-data policy. More seriously, the released pipeline accepts a row whenever sentiment is valid; it does not require the other fields or enforce tag vocabulary membership. The 59,708 accepted rows contain 290 empty justifications, 32 empty captions, 13 empty tag lists, and 6,820 out-of-vocabulary selections affecting 6,652 rows. Only 53,043 rows have a non-empty, fully valid tag list. Some records leak persona commentary into captions while leaving justification empty. The paper does not describe cleaning for the exact fields it analyzes. The 50 nominal agents per profile are not independent persons or prompts: persona ID is absent from the prompt, so they are repeated stochastic calls with identical labels, image, model, seed, and configuration. Text varies, but the generalization units remain 24 prompts, 50 images, and one model. Construct interpretation is further limited because the prompt mandates persona-independent captions and persona-voiced justifications. Thus the central contrast is partly an instruction manipulation check. Released examples literally repeat conservative, low-income, and progressive framing, so measured separation may be label leakage and stereotype reproduction rather than human perception. Pragmatic, Empathetic, and Analytical are not a validated personality instrument, and no demographically matched human outputs establish persona validity. The per-image within/cross summaries are naturally paired, but the paper reports unpaired Mann-Whitney tests. At least twelve comparisons lack multiplicity correction, confidence intervals, or standardized effects. Profile-pair Pearson p-values ignore shared profiles and images and require matrix-aware permutation. BERTopic configuration, outlier handling, topic naming, Jensen-Shannon values, tests, and uncertainty are absent; topic patterns are also strongly confounded with sentiment. The no-persona comparison contrasts one realization and a centroid with average pairwise persona similarity, so it cannot rule out stochastic variation despite the paper's claim. Finally, the manuscript promises a public repository, but the rendered note says it will appear after publication and the commented GitHub URL is unavailable. The upstream corpus repository exposes generation and raw outputs, enabling this integrity audit, but not the embeddings, topic model, statistical code, or processed artifacts for this paper. The release supports prompt-conditioned linguistic variation, not faithful demographic simulation, human perceptual validity, improved annotation, psychological causality, or generalization across models, cultures, images, and prompts.
Research question
Do persona labels affect the descriptive grounding, the perception labels, or the interpretive framing of an MLLM more when evaluating the same urban scenes?