Urban region representation learning plays a fundamental role in urban computing, providing compact embeddings that support a wide range of downstream analytical tasks. Existing methods are developed under a single-city setting and rely on continuous vector representations learned from city-specific spatiotemporal data. Learning transferable and semantically meaningful region representations in cross-city scenarios remains a significant challenge. In this paper, we propose URBAN-PROTOS, a two-stage framework for cross-city learning of urban region representations. We first abstract multi-city mobility data into discrete Urban Prototypes, which capture city-invariant spatiotemporal regularities from localized heterogeneous contexts. These prototypes serve as shared semantic units across cities and are further integrated into the semantic space of large language models through a prototype-guided alignment mechanism. Extensive experiments conducted demonstrate that URBAN-PROTOS consistently outperforms existing methods in cross-city transfer settings, while maintaining strong performance in single-city evaluations. Further analyses show that the learned region embeddings encode deep semantic regularities shared across cities, providing robust support for cross-city generalization.

Large reasoning models (LRMs) achieve strong performance through extended reasoning traces, but they often exhibit overthinking behavior for low-complexity queries. Existing efforts to mitigate this issue are fundamentally limited by unstable accuracy–efficiency trade-offs and poor robustness to heterogeneous reasoning behaviors. To address these challenges, we propose a two-stage framework for stable adaptive thinking in LRMs. The framework first applies Hybrid Fine-Tuning to expose the model to both thinking and no-thinking behaviors, establishing well-conditioned initialization. It then performs adaptive reinforcement learning with Correctness-Preserving Advantage Shaping (CPAS) to avoid suppressing correct long-chain reasoning, and Length-Aware Gradient Regulation (LAGR) to stabilize optimization under severe reasoning-length heterogeneity. Extensive experiments on Qwen2.5-1.5B and 7B show consistent improvements over strong baselines, achieving up to +3.7/+3.6 accuracy points while reducing generated tokens by 40.6%/43.9%. Further analyses across varying problem difficulties and out-of-distribution tasks confirm the robustness and generalization of our approach.