Generative agents offer promising capabilities for simulating realistic urban behaviors. However, existing methods often rely on static profiles, oversimplified behavioral logic, and synchronous inference pipelines that hinder scalability. We present MobileCity, a lightweight generative-agent framework for city-scale simulation powered by cognitively-grounded generative agents. Each agent acts based on its needs, habits, and obligations, evolving over time. Agents are initialized from survey-based demographic data and navigate a realistic multimodal transportation network spanning multiple types of vehicles. To achieve scalability, we introduce asynchronous batched LLM inference during action selection and a low-token communication mechanism. Experiments with 4,000 agents demonstrate that MobileCity generates more human-like urban dynamics than baselines while maintaining high computational efficiency. Our code is publicly available at https://github.com/Tony-Yip/MobileCity.

With the rapid growth of online counseling services, timely and reliable risk classification of counseling records is essential for supporting early screening and prioritizing limited intervention resources. High-risk samples refer to high-acuity suicide risk and require expedited human review. However, this task is challenging due to severe class imbalance (93% low-risk and 7% high-risk samples) and complex decision boundaries. Large language models (LLMs) exhibit unstable predictions and systematic errors in such imbalanced clinical-text settings. To address this issue, we propose Disagreement-Driven Joint Refinement (DDJR), an iterative, parameter-free refinement framework. It uses prediction disagreement between two inference settings, zero-shot and retrieval-augmented in-context learning, as the primary signal for identifying high-value instances. These disagreement-identified instances are transformed into adaptive refinement signals and used to jointly update both the exemplar pool and an executable rule set, thereby sharpening decision boundaries and improving prediction stability. Experiments on 6,481 real-world counseling records demonstrate that the proposed DDJR outperforms existing methods, achieving an accuracy of 0.915 and a Matthews Correlation Coefficient (MCC) of 0.583. These results demonstrate that DDJR achieves more stable and reliable predictions for high-stakes counseling risk classification in real-world settings.