Unsupervised Domain Adaptation (UDA) aims to mitigate distribution shifts between labeled source domains and unlabeled target domains for robust cross-domain decision-making. A common strategy is to employ pseudo-labeling to approximate target supervision, but this often introduces significant label noise due to domain discrepancies, severely limiting model generalization. We identify the root cause as the scarcity of high-confidence pseudo-labeled target domain samples (hcpl-TDS), which impairs reliable statistical modeling in the target domain. To address this challenge, we propose a latent space optimization framework based on flow matching, which enables controllable generation of diverse and semantically consistent hcpl-TDS. This process is guided by a class-conditional anchor mechanism that captures the conditional structure of both source and target domains, ensuring semantic alignment in the generated features. Leveraging these synthesized hcpl-TDS, we further enable reliable conditional alignment to enhance decision boundary optimization under distribution shift. Extensive experiments on standard UDA benchmarks validate the effectiveness of our method, consistently outperforming state-of-the-art approaches. Detailed ablation studies provide further insight into the contributions of each component.