Overview
This study examines the influence of persona prompting on language generated by two multimodal large language models (LLMs) in the context of urban perception. Urban perception is defined as a setting where subjective interpretations of shared visual evidence are investigated. The research organized LLM outputs into three functional levels: descriptive grounding (captions), an intermediate semantic layer (perception tags), and interpretive framing (justifications).
Research Context
The study specifically focused on how persona prompting shapes the outputs of multimodal LLMs. The chosen application domain, urban perception, was utilized as a framework to explore subjective interpretations derived from common visual data. This context underscores the interest in understanding how pre-defined personas impact the nuanced understanding and description of visual scenes by AI models.
Approach
The research employed two specific multimodal large language models: Qwen3-VL-8B and Gemma-4-E4B-it. For each model, approximately 60,000 persona-conditioned annotations were generated. These annotations were analyzed across three distinct functional output levels: captions, perception tags, and justifications. The methodology involved evaluating the convergence and variation of these outputs based on different persona profiles, with specific attention to attributes such as economic status, political orientation, and personality.
Findings
- Descriptive Grounding (Captions): Captions generated by the models exhibited strong convergence across various persona profiles. Differences observed in captions were generally small and specifically associated with particular attributes.
- Interpretive Framing (Justifications): Justifications showed substantially greater variation compared to captions. Economic status emerged as the attribute producing the largest difference in justifications across both models. Political orientation and personality were also identified as prominent factors contributing to significant variation in justifications.
- Perception Tags: For perception tags, personas sharing the same attribute level were found to produce more similar tag sets than personas characterized by different attribute levels. The most significant separation in perception tag similarity was observed for economic status.
- Image-Level Comparisons: Paired image-level comparisons reinforced the finding that justification differences were larger than caption differences when considering economic status, political orientation, and personality attributes.
- Exploratory Topic Analysis: An exploratory topic analysis further indicated the presence of persona-specific evaluative emphasis within the generated outputs.
- Cross-Model Consistency: Across both Qwen3-VL-8B and Gemma-4-E4B-it, the patterns of profile-pair similarity were strongly correlated for all three output types (captions, perception tags, and justifications). However, the lowest agreement between the two models was observed for justifications.
- Overall Impact of Persona Prompting: The study concluded that persona prompting exerts a stronger effect on interpretive framing than on descriptive grounding in the context of multimodal urban perception.
Why This Matters
The findings indicate that while multimodal large language models provide largely consistent descriptive grounding of visual information regardless of persona, the subjective and interpretive aspects of their outputs are significantly influenced by specified persona attributes. This distinction is relevant for applications requiring nuanced understanding or perspective-taking from AI, as it highlights areas where persona prompting can effectively modulate interpretive framing. The observed correlation in profile-pair similarity across models suggests a consistent underlying mechanism for persona integration, despite varying levels of agreement on specific output types.