Generative Recommendation (GR) has demonstrated remarkable performance in next-token prediction paradigms, which relies on Semantic IDs (SIDs) to compress trillion-scale data into learnable vocabulary sequences. However, existing methods suffer from three critical limitations: (1) Information Degradation: the two-stage compression pipeline causes semantic loss and information degradation, with no posterior mechanism to distinguish high-quality from low-quality SIDs; (2) Semantic Degradation: cascaded quantization discards key semantic information from original multimodal features, as the embedding generation and quantization stages are not jointly optimized toward a unified objective; (3) Modality Distortion: quantizers fail to properly align text and image modalities, causing feature misalignment even when upstream networks have aligned them. To address these challenges, we propose a novel framework integrating three key innovations: Deep Contextual Interest Mining (DCIM), Cross-Modal Semantic Alignment (CMSA), and Quality-Aware Reinforcement Mechanism (QARM). First, we leverage Vision-Language Models (VLMs) to align non-textual modalities into a unified text-based semantic space, mitigating modality distortion. Second, we introduce a deep interest mining mechanism that captures high-level semantic information implicitly present in advertising contexts, encouraging SIDs to preserve critical contextual information through reconstruction-based supervision. Third, we employ a reinforcement learning framework with quality-aware rewards to encourage semantically rich SIDs while suppressing low-quality ones in the posterior stage. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art SID generation methods, achieving superior performance on multiple benchmarks. Ablation studies further validate the effectiveness of each proposed component

翻译：暂无翻译