This paper proposes HiD-VAE, a novel framework for learning hierarchically disjoint item representations, to address the semantic flatness and representational entanglement problems caused by unsupervised tokenization in existing generative recommender systems. HiD-VAE aligns multi-level item tags and discrete codes through a hierarchical supervised quantization process to generate more uniform and disjoint IDs, and introduces uniqueness loss to directly penalize latent space redundancy, thereby addressing the representational entanglement problem. This approach achieves improved recommendation accuracy and diversity, and we demonstrate its superior performance over state-of-the-art methods through experiments on three publicly available benchmarks.
