CityRiSE: Reasoning Urban Socio-Economic Status in Large Vision-Language Models via Reinforcement Learning

• Existing methods have limited transfer to unseen cities or indicators, require substantial labeled data, and often provide no interpretable reasoning.LVLM-based approaches may rely on tens of thousands of supervised samples.Source: §1 · p. 2
• CityRiSE normalizes diverse indicators into a unified prediction format and uses reinforcement learning to learn transferable reasoning patterns from visual inputs.The approach combines GRPO with auxiliary perceptual and general visual reasoning datasets and reports using 5,109 training samples.Source: §1 · p. 2
• The paper reports cross-city and cross-indicator generalization, including the first demonstrated cross-indicator generalization in this domain.The authors describe this as emergent visual reasoning enabled by reinforcement learning with LVLMs.Source: §1 · p. 2
• CityRiSE is reported to outperform state-of-the-art baselines across 10 urban socio-economic prediction tasks, especially on unseen cities and indicators.The authors also report performance exceeding proprietary commercial LVLMs despite using substantially fewer training samples.Source: §1 · p. 2

2.1 Transferable Datasets for Cross-City and Cross-Indicator Generalization

• The dataset suite combines socio-economic prediction data with two auxiliary datasets targeting transferable perceptual and logical reasoning skills.The CityLens dataset covers 17 cities; indicators are discretized into 10 bins, and the evaluation includes held-out cities and indicators.Source: §2.1 · p. 2, §2.1.1 · p. 2

2.1.2 Perceptual Urban Reasoning Data.

• Perceptual tasks use training-split images and annotations to build urban reasoning skills, including spatial layout, geographic context, and socio-economic comparison.Spatial reasoning distinguishes a distant image from two nearby ones; geolocation predicts image city of origin, while ranking compares indicator levels between regions.Source: §2.1.2 · pp. 2–3

2.1.3 General Visual Reasoning Data.

• General visual reasoning trains object counting and pattern completion, encouraging transferable skills in quantifying objects and inferring visual regularities.Pattern completion presents a sequence or matrix with a missing element and asks the model to infer the rule and select an option.Source: §2.1.3 · p. 3

2.2 Guided Reward Design for Interpretable Socio-Economic Reasoning

• The dual reward design pairs a keyword signal for goal-directed, interpretable reasoning with a regression signal for numerical alignment and objective evaluation.Table 2 summarizes reward schemes across training tasks and instance statistics.Source: §2.2 · p. 3, Table 2 · p. 4

2.2.1 Keyword Reward.

• The Keyword Reward encourages coverage of task-relevant visual evidence without prescribing a fixed reasoning template.Its six urban-perceptual keywords are person, vehicle, greenery, road infrastructure, street furniture, and building; location is also specially rewarded as a geographic anchor.Source: §2.2.1 · p. 4

2.2.2 Regression Reward.

• The Regression Reward applies Huber loss so small prediction errors are penalized less harshly than large ones.Its exponential form yields smoothly decaying rewards, with Huber loss behaving quadratically near correct values and linearly for larger errors.Source: §2.2.2 · p. 4

• For each data point, GRPO generates multiple candidate responses and calculates a reward for each using the paper’s reward design.The algorithm takes a pretrained policy, reference policy, reward model, and training dataset as inputs.Source: §2.3 · p. 4
• GRPO computes a group-normalized advantage for each response, reflecting its reward relative to other candidates in the same group.Source: §2.3 · p. 4
• The policy is updated by minimizing a loss that incorporates the advantage and KL divergence, encouraging performance improvement while accounting for divergence from the reference policy.The algorithm returns the optimized policy.Source: §2.3 · p. 4

3.2 Overall Performance (RQ1)

• 0.200 R² on Life Expectancy and 0.218 R² on House Price accompany CityRiSE’s best overall performance across most unseen indicators.CityRiSE maintains positive R² scores on all unseen indicators.Source: §3.2 · p. 5, Table 3 · p. 5
• 0.286 R² on Bachelor Ratio is CityRiSE’s highest performance on unseen cities, while its overall spatial-transfer trend remains consistent across tasks.The authors note weaker Public Transport performance and suggest unclear visual cues and relatively low test-city values may partially limit generalization.Source: §3.2 · p. 5, Table 3 · p. 5
• −0.330 to 0.334 on Population and −0.013 to 0.594 on Public Transport are CityRiSE’s R² improvements over Qwen2.5-VL-7B in-domain.SFT models lead on some indicators, particularly GDP and Population; comparable SVF-GT and SVF-Rank results suggest normalization preserves predictive accuracy, without implying full equivalence.Source: §3.2 · pp. 5–6, Table 3 · p. 5

3.3 Ablation Study (RQ2)

• Removing either reward degrades performance, and removing both causes severe failure; Regression Reward supports numerical alignment while Keyword Reward strengthens semantic grounding.With both removed, GDP and Public Transport R² fall from 0.39 and 0.59 to below −1.00; removing Keyword Reward drops Violent Crime from 0.30 to −0.14, while removing Regression Reward drops unseen-city Population from 0.36 to 0.04.Source: §3.3 · p. 6, Figure 2 · p. 6
• Removing either auxiliary dataset degrades performance across most indicators, with General Visual Reasoning data broadly influential and Perceptual Urban Reasoning data more targeted.Removing General Visual Reasoning data lowers Mental Health R² from 0.26 to −0.71; removing Perceptual Urban Reasoning data notably affects Life Expectancy and House Price, and removing both causes severe degradation.Source: §3.3 · p. 6, Figure 3 · p. 6

3.4 Training Design Validation and Input Modality Analysis (RQ3)

• CityRiSE outperforms its generic-keyword variant on all six tasks, while adding auxiliary data to SFT yields modest in-domain gains but declines on most out-of-domain tasks.The comparison tests task-aligned visual keywords against generic thinking tokens; the authors conclude auxiliary data alone is insufficient to produce CityRiSE’s transferable gains.Source: Figure 4 · p. 6, §3.4 · p. 7
• Removing satellite imagery causes moderate declines, whereas removing street-view imagery causes larger drops, including negative R² on House Price, Mental Health, and Public Transport.The authors characterize street-view imagery as the primary fine-grained evidence and satellite imagery as complementary regional-scale context.Source: §3.4 · p. 7, Table 4 · p. 7

3.5 Reasoning Analysis and Comparison (RQ4)

• The no-reasoning variant beats the base model but remains below CityRiSE across all three evaluation settings; comparisons also report more flexible transfer from reward-induced reasoning than scripted CoT.In the qualitative comparison, CityRiSE produces coherent reasoning without human-crafted chains; later RL on UrbanMLLM does not induce reasoning, which the authors suggest may reflect rigidity from omitting reasoning at the early SFT stage.Source: §3.5 · p. 7

Socio-Economic Prediction from Urban Imagery

• Urban socio-economic sensing predicts high-level indicators from urban visual data, using approaches from classic regression to deep learning and representation learning.LLMs have also been used to add contextual information or directly predict indicators.Source: §4 · pp. 7–8

Generalist Reasoning in LVLMs

• Figure 5 contrasts the reasoning patterns of CityRiSE and baseline models, while Figure 6 compares CityRiSE variants and training methods across representative tasks and three task types.The captions describe Figure 5 as a qualitative comparison and Figure 6 as a performance comparison; they do not state specific outcomes for Figure 6.Source: Figure 5 · p. 8, Figure 6 · p. 8

Socio-Economic Prediction from Urban Imagery

• GeoLLM links an LLM’s internal knowledge to indicators through address information, while UrbanMLLM uses large-scale pretraining and supervised learning for multi-type, multi-task prediction.Other paradigms use LLMs as supplementary tools or direct predictors.Source: §4 · p. 8

Generalist Reasoning in LVLMs

• Prior LVLM reasoning work mainly targets intuitive, low-level visual elements, including geometry problems and natural-image comprehension.Such work has used reasoning-annotated datasets and designed reward functions to guide visual reasoning.Source: §4 · p. 8
• CityRiSE applies reinforcement learning to implicit, high-level visual reasoning, associating intuitive visual elements with hidden knowledge for generalization to unseen indicators and cities.The passage identifies this as the first such application of RL to high-level visual reasoning.Source: §4 · p. 8

5 Conclusion

• CityRiSE outperforms state-of-the-art baselines on 10 urban socio-economic prediction tasks.The conclusion presents it as the first framework for generalist prediction across diverse cities and indicators using emergent LVLM reasoning through reinforcement learning.Source: §5 · p. 8
• Normalized prediction targets encourage transferable reasoning patterns connecting visual inputs to high-level socio-economic concepts, supporting cross-city and cross-indicator generalization.The conclusion describes generalization across diverse cities and indicators as the framework’s scope.Source: §5 · p. 8
• GRPO and two auxiliary datasets for transferable perceptual and logical reasoning support strong performance with 5k training samples.The conclusion characterizes this as high data efficiency.Source: §5 · p. 8
• A verifiable multi-component reward mechanism induces coherent, goal-directed reasoning and enhances interpretability and prediction quality.Source: §5 · p. 8