Evaluating 15-minute life circles at block scale with multimodal LLMs: Where they work, where they fail, and what to do next

The 15-min life circle (15-LC) has become an adopted paradigm for urban governance and public service improvement, yet block-scale evaluation remains fragmented, with a gap between indicator computation, integrated interpretation, and actionable translation. Leveraging recent advances in multimodal large language models (MLLMs), we propose a 5D integrated framework for policy-defined 15-LC units and operationalize it as a protocolized benchmark with standardized semantic-map inputs and structured outputs. We benchmark nine MLLMs across providers, generations, and reasoning modes, and apply the framework to 216 inner-ring 15-LC units in Shanghai. Quantitatively, the best model, Gemini-3 (non-thinking), achieves an overall tolerance accuracy of 0.682; counting and distribution-summary tasks are most robust (TR = 0.73), whereas scale-sensitive measurements remain the main bottleneck (TR = 0.48–0.51). Newer-generation models improve overall performance (TR = 0.59 vs. 0.54), but Thinking does not yield consistent gains in quantitative accuracy. Qualitatively, model ratings show positive rank agreement with expert assessments and generate evidence-grounded, spatially targeted recommendations, though scores are more compressed with fewer extremes. Empirically, Shanghai's inner-ring 15-LCs exhibit a core–periphery structure: the core performs better in functional diversity and accessibility, while peripheral units more often face longer service distances and weaker transit connections; system-wide weaknesses concentrate in street-network design (lowest mean score, 3.63 ± 0.64) and density suitability shows low-value clustering in the core, suggesting renewal should shift from facility addition toward micro-connectivity and density balancing. Overall, the benchmark clarifies a human–AI workflow in which MLLMs support evidence integration, interpretable diagnosis, and recommendation drafting, while GIS and experts anchor scale-sensitive measurement and contextual validation.