Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Urban-R1: Reinforced MLLMs Mitigate Geospatial Biases for Urban General Intelligence

Qiongyan Wang, Xingchen Zou, Yutian Jiang, Haomin Wen, Jiaheng Wei, Qingsong Wen, Yuxuan Liang

TL;DR

Urban-R1 tackles geospatial bias in urban foundation models trained with supervised fine-tuning by introducing a reinforcement learning post-training framework that uses Group Relative Policy Optimization (GRPO) and an Urban Region Profiling (URP) proxy task. By fusing multimodal inputs (satellite imagery, coordinates, and text) and rewarding geography-grounded, verifiable reasoning, Urban-R1 learns invariant regional patterns and improves cross-region generalization. Across seen and unseen regions, it achieves superior geo-bias mitigation and strong cross-task performance, outperforming SFT baselines and rivaling closed-source models while remaining open-source. The work highlights reinforcement learning alignment as a practical path toward equitable, trustworthy urban intelligence and points to future extensions to tool use and real-time analytics for complex urban decision-making.

Abstract

Rapid urbanization intensifies the demand for Urban General Intelligence (UGI), referring to AI systems that can understand and reason about complex urban environments. Recent studies have built urban foundation models using supervised fine-tuning (SFT) of LLMs and MLLMs, yet these models exhibit persistent geospatial bias, producing regionally skewed predictions and limited generalization. To this end, we propose Urban-R1, a reinforcement learning-based post-training framework that aligns MLLMs with the objectives of UGI. Urban-R1 adopts Group Relative Policy Optimization (GRPO) to optimize reasoning across geographic groups and employs urban region profiling as a proxy task to provide measurable rewards from multimodal urban data. Extensive experiments across diverse regions and tasks show that Urban-R1 effectively mitigates geo-bias and improves cross-region generalization, outperforming both SFT-trained and closed-source models. Our results highlight reinforcement learning alignment as a promising pathway toward equitable and trustworthy urban intelligence.

Urban-R1: Reinforced MLLMs Mitigate Geospatial Biases for Urban General Intelligence

TL;DR

Urban-R1 tackles geospatial bias in urban foundation models trained with supervised fine-tuning by introducing a reinforcement learning post-training framework that uses Group Relative Policy Optimization (GRPO) and an Urban Region Profiling (URP) proxy task. By fusing multimodal inputs (satellite imagery, coordinates, and text) and rewarding geography-grounded, verifiable reasoning, Urban-R1 learns invariant regional patterns and improves cross-region generalization. Across seen and unseen regions, it achieves superior geo-bias mitigation and strong cross-task performance, outperforming SFT baselines and rivaling closed-source models while remaining open-source. The work highlights reinforcement learning alignment as a practical path toward equitable, trustworthy urban intelligence and points to future extensions to tool use and real-time analytics for complex urban decision-making.

Abstract

Rapid urbanization intensifies the demand for Urban General Intelligence (UGI), referring to AI systems that can understand and reason about complex urban environments. Recent studies have built urban foundation models using supervised fine-tuning (SFT) of LLMs and MLLMs, yet these models exhibit persistent geospatial bias, producing regionally skewed predictions and limited generalization. To this end, we propose Urban-R1, a reinforcement learning-based post-training framework that aligns MLLMs with the objectives of UGI. Urban-R1 adopts Group Relative Policy Optimization (GRPO) to optimize reasoning across geographic groups and employs urban region profiling as a proxy task to provide measurable rewards from multimodal urban data. Extensive experiments across diverse regions and tasks show that Urban-R1 effectively mitigates geo-bias and improves cross-region generalization, outperforming both SFT-trained and closed-source models. Our results highlight reinforcement learning alignment as a promising pathway toward equitable and trustworthy urban intelligence.
Paper Structure (23 sections, 5 equations, 9 figures, 6 tables)

This paper contains 23 sections, 5 equations, 9 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Urban General Intelligence
  4. Reinforcement Learning for Large Models
  5. Methodology
  6. Urban Region Profiling as a Proxy Task
  7. Reinforcement Learning for URP
  8. Reducing Geo-Bias: RL vs. SFT
  9. Enhancing Reasoning via GRPO
  10. Comparison to Existing Arts
  11. Experiments
  12. Experimental Settings
  13. Mitigating Geo-bias (RQ1)
  14. Superior Performance on Seen Regions
  15. Results on Unseen Regions
  16. ...and 8 more sections

Figures (9)

  • Figure 1: An example of geographic bias in regional GDP prediction by GPT-4o. The top panel depicts the ground truth of GDP rankings (color-coded: green indicates a higher rank, red indicates a lower rank), while the bottom panel shows predictions from GPT-4o.
  • Figure 2: Training pipeline of Urban-R1.
  • Figure 3: Comparison of house price inference by SFT and RL models, showing RL's more accurate reasoning aligned with the correct answer.
  • Figure 4: Geographic distribution of training (blue) and unseen test (orange) regions.
  • Figure 5: Poverty prediction ranks on unseen regions, with indicating higher ranks and red representing lower ranks. Notably, Urban-R1 aligns more closely with the ground truth pattern compared to the SFT baseline.
  • ...and 4 more figures