Stochastic Streets: Biases and Limitations of LLMs in Random Address Generation

23/02/2026

In “Stochastic Streets: A Walk Through Random LLM Address Generation in Four European Cities,” published in IEEE Computer Society’s journal Computer, researchers from the Information Processing and Telecommunications Center (IPTC) at Universidad Politécnica de Madrid (UPM), together with collaborators from Politecnico di Milano, Telefónica Tech, FIWARE Foundation, and Northeastern University, investigate whether large language models (LLMs) can generate truly random street addresses.

The team evaluated six state-of-the-art models across Amsterdam, Madrid, Paris, and Rome, issuing 1,000 prompts per city and model. Despite explicit instructions to ensure uniform randomness, results show strong biases: models repeatedly selected a small subset of streets (sometimes hundreds of times) while ignoring thousands of alternatives. In some cases, they even generated non-existent streets. Address numbers were generally more evenly distributed than street names, yet clear numerical preferences (e.g., 12, 14, 123) emerged, especially in formats where the number appears first.

Geospatial heatmaps revealed that certain models concentrated outputs in just a few locations, exposing systematic structural biases rather than true stochastic behavior. These findings highlight limitations in how LLMs internally represent geographic and structured data, and how token prediction mechanisms influence supposedly “random” outputs.

The research has practical implications for logistics simulation, traffic optimization, emergency response planning, tourism route generation, synthetic data creation, geocoding validation, and bias analysis in AI systems. Future work will explore fine-tuning strategies to improve randomness and generalization across cities.

Bibliographic reference:

Fu, T., Nazareno-Campo, D., Coronado-Blázquez, J., Conde, J., Reviriego, P., & Lombardi, F. Stochastic Streets: A Walk Through Random LLM Address Generation in Four European Cities. Computer, 59 (3), pp. 92 – 99. https://doi.org/10.1109/MC.2025.3611067

Javier Conde, GS / ORCID / LinkedIn

Pedro Reviriego, GS / ORCID / LinkedIn

LinkedIn: https://www.linkedin.com/company/iptc-upm/

For more information: www.iptc.upm.es