Vision-Language Models in Urban Planning: Progress and Pitfalls

Spatial planning maps play a important role in territorial governance by transforming planning objectives and regulations into visual formats for better decision-making and public communication. However, interpreting these maps requires not only acute visual perception but also nuanced spatial reasoning and professional judgment informed by policy. This complexity poses significant challenges for both human learners and AI systems. With the rapid development of Vision-Language Models (VLMs), their application in urban planning analysis is gaining traction, yet existing benchmarks are predominantly focused on general visual comprehension, often neglecting the domain-specific cognitive processes key in planning.

The Introduction of PlanBench-V

To bridge this gap, the launch of PlanBench-V marks a significant advancement. This comprehensive benchmark is designed to assess the effectiveness of VLMs in spatial planning map interpretation. Central to this effort is the Spatial Planning Map Database (SPMD), an expertly annotated collection of 223 planning maps and 1,629 question-answer pairs curated by professional planners. Covering diverse geographic regions and cartographic styles, the SPMD provides a solid foundation for evaluation.

An evaluation framework was devised, informed by planning theory, to assess four key capabilities: Perception, Reasoning, Association, and Implementation. These are intended to mirror the cognitive sequence necessary for interpreting planning maps. Initial experiments with two generations of VLMs reveal clear progress, but also persistent challenges. The 2026 model, Qwen3.6-Plus, showed a notable improvement by outperforming its predecessor, GPT-4o, by 27%. Despite this advancement, all models continue to struggle with tasks requiring evaluative judgment, policy sensitivity, and constraint-aware decision-making.

AI's Limitations in Professional Planning

These findings bring to light the fundamental limitations of current VLMs in professional planning contexts. While the progress is commendable, it's evident that these models aren't yet equipped to handle the intricacies of implementation-oriented tasks within urban planning. The question that arises is, can AI ever fully grasp the multifaceted nature of urban governance? While the technology continues to evolve, its current shortcomings suggest a need for more domain-adaptive multimodal reasoning frameworks.

For institutional allocators and those managing diversified portfolios, the implications of these advancements are significant. Investing in AI technologies for urban planning presents opportunities, yet demands caution. While the risk-adjusted case remains intact, position sizing warrants careful review. The custody question remains the gating factor for most allocators when considering tangible applications of AI in their portfolios.

The Path Forward

As we look to the future, the development of VLMs for urban planning must prioritize domain-specific capabilities. The current trajectory shows promise, yet the journey to creating AI models that can truly understand and implement planning principles is far from over. Fiduciary obligations demand more than conviction. They demand a process. Before discussing returns, we should discuss the liquidity profile. Will AI's evolution in this sector match the pace of its investment potential? Only time, and continued innovation, will tell.