<p>Generative recommendation leverages the strong semantic understanding and contextual reasoning capabilities of Large Language Models (LLMs), enabling improved handling of long-standing challenges such as cold-start, explainability, and conversational interaction. However, existing generative and sequential recommendation methods still suffer from key limitations, including insufficient user representation learning, inadequate sequence modeling, and weak alignment between collaborative signals and textual semantics. This work identifies three central issues: (1) Behavior sequence sparsity, which restricts the ability to capture comprehensive user interest patterns. (2) Structural homogeneity, where simple sequential structures fail to reflect diverse behavioral relationships. (3) Semantic inconsistency between collaborative-filtering embeddings and textual representations, weakening cross-feature fusion. To address these challenges, we propose an enhanced alignment modeling framework integrating sequence augmentation, Mixup fusion, and contrastive semantic alignment. Two structural augmentation strategies, M-Reorder and M-Substitute, generate multi-view behavioral sequences that uncover latent behavioral patterns and enrich user representations. A Mixup strategy then fuses embeddings from different augmented views, stabilizing user preference modeling and producing more robust representations. Furthermore, an InfoNCE contrastive loss is introduced to explicitly enhance the consistency between item embeddings and textual semantics within a shared embedding space. Experiments on multiple public datasets demonstrate that the proposed framework improves overall performance by 1.5%–2.5%, confirming its effectiveness in addressing sparsity, structural simplicity, and semantic discrepancy while enhancing both accuracy and generalization.</p>