Exploring multimodal large language models’ potential in simulating human perception of urban cycling environments: A street-view perspective

Urban cycling is a key component of sustainable transportation, and its development depends on creating cyclist-friendly environments. However, traditional data collection, including surveys and street view imagery (SVI) annotations, is constrained by limited scale and high cost, creating a bottleneck for systematic, city-level analysis. The development of Multimodal Large Language Models (MLLMs) offers new possibilities, but recent studies often focus on numerical scores while neglecting the explanatory textual rationales produced. This study first validates the MLLM's performance against a human-rated dataset, then develops a Cascaded Perception Parsing Framework (CPPF) to explain its decision-making behavior. The framework features a two-stage process: a secondary Large Language Model (LLM) first parses the MLLM's textual rationales into structured data, which then informs an interpretable XGBoost proxy model for analysis based on eXplainable Artificial Intelligence (XAI). First, the MLLM shows high agreement with human ratings for Traffic Safety and Public Security but struggles with the more subjective Scenic Beauty. Second, the proxy model analysis attributes this discrepancy to the MLLM's simplistic aesthetic logic: a mechanistic preference for orderliness that ignores complex human intuition. This research provides critical empirical evidence on the capability boundaries and cognitive limitations of AI in urban perception, while also offering a novel methodology for developing more explainable AI-assisted tools.