Autonomous Urban Region Representation with LLM-informed Reinforcement Learning

ICLR 2026 Conference Submission10751 Authors

Keywords: Urban Representation Learning, Reinforcement Learning, Large Language Models

TL;DR: We propose SubUrban, a submodular-aware RL framework that leverages LLM guidance for automated urban region representation learning.

Abstract: Urban representation learning has become a key approach for many applications in urban computing, but existing methods still rely heavily on manual feature designs and geographic heuristics. We present SubUrban, a reinforcement learning framework that autonomously discovers informative regional features through submodular rewards and semantic guidance from large language models. SubUrban adaptively expands each region into a hypernode, suppressing redundancy while preserving complementary associations, and learns cross-task embeddings with a graph-attention policy. Experiments across multiple prediction tasks (population, house price, and GDP) and cities (Beijing, Shanghai, New York, and Singapore) show that SubUrban consistently outperforms state-of-the-art baselines, achieving comparable accuracy with only 10\% of the training data. These results highlight submodular-driven automation, enhanced by LLM-in-the-loop semantics, as a practical paradigm for autonomous urban region representation learning. The implementation of our SubUrban is available at \url{https://anonymous.4open.science/r/SubUrban_ICLR2026}.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 10751