Urban region representation learning aims to embed heterogeneous signals such as mobility flows, points of interest, and land-use patterns into unified vectors for downstream tasks like check‑in prediction, crime forecasting, and service demand estimation.\ \ Existing approaches rely on a globally shared fusion layer and joint optimization, implicitly assuming uniform modality importance and naturally balanced multi‑task training.\ \ In practice urban data is highly heterogeneous. Our empirical study reveals two structural issues:\ (i) Modality utility varies dramatically across regions, causing static fusion to over‑emphasize certain signals;\ (ii) Gradient magnitudes across objectives remain imbalanced, leading to representational bias and unstable training.\ \ We introduce PRISM, a self‑adjusting framework for adaptive multi‑view learning. PRISM employs sparse MoE routing to allocate views on a per‑region basis, allowing fusion weights to adapt to local characteristics. At the same time, a PID‑inspired regulation mechanism monitors gradient conflicts and mitigates them, balancing the learning rates of all tasks.\ \ Experiments on three large‑scale urban datasets show that PRISM achieves performance comparable to or slightly better than strong multi‑view baselines, while markedly improving optimization stability and producing more balanced region representations.\ \ Review: By combining region‑aware routing with gradient regulation, PRISM breaks the rigidity of global fusion and offers a scalable solution for joint learning from diverse urban data sources.