Do LLMs Know Your Neighborhood? Auditing LLM Priors for Neighborhood-Level Mobility Prediction and Structural Alignment

Research is evaluating the capability of Large Language Models (LLMs) to predict neighborhood-level human mobility, which is crucial for urban planning, transportation, public health, and emergency response. This study investigates whether zero-shot LLMs can infer aggregate mobility patterns at the Census Block Group level across U.S. metropolitan areas, comparing their performance against supervised baselines. This addresses the challenge of accessing fine-grained, privacy-sensitive trajectory data by exploring LLMs as a potential alternative.

Why it matters

This research addresses a critical need for accessible, granular human mobility data, which is foundational for effective urban infrastructure development, public safety, and resource allocation. If LLMs prove capable, they could offer a scalable and less resource-intensive alternative to proprietary data, significantly enhancing strategic planning and operational responsiveness across numerous public and private sectors.

Key insights

• Human mobility data is vital for critical sectors like urban planning, transportation, public health, and emergency response.
• Fine-grained human trajectory data is typically proprietary, restricted, and sensitive to privacy concerns.
• Large Language Models (LLMs) are being explored as a potential method to generate plausible mobility traces and predict individual movement.
• The research aims to clarify if LLMs can effectively infer aggregate neighborhood-level mobility, as their capacity for this is currently uncertain.
• Zero-shot LLMs are being evaluated for mobility prediction at the Census Block Group level across four U.S. metropolitan areas.
• The evaluation uses anonymized Cuebiq data to establish various mobility outcomes (point-level, trajectory-level, temporal) and links them with sociodemographic and built-environment predictors.
• LLM predictions are being compared against established supervised baselines.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.00345