IDGenRec: LLM-RecSys Alignment with Textual ID Learning This paper proposes IDGenRec, a novel framework for aligning large language models (LLMs) with recommendation systems by learning textual IDs for items. The framework generates unique, concise, and semantically rich textual IDs using human language tokens, enabling seamless integration of personalized recommendations into natural language generation. The ID generator is trained alongside the LLM-based recommender, allowing for collaborative learning and improved recommendation performance. The ID generator produces item IDs based on item metadata, ensuring that the IDs are short, unique, and meaningful for recommendation purposes. A diverse ID generation algorithm is employed to ensure uniqueness, and the generated IDs are integrated into the text-to-text generation process. The base recommender then generates the target item ID based on the generated IDs and user history. The framework is evaluated on four standard sequential recommendation datasets and shows significant improvements over existing baselines. Additionally, the framework is tested in a zero-shot setting on six unseen datasets, demonstrating strong performance and the potential of IDGenRec as a foundation model for generative recommendation. The results show that the zero-shot performance of the pre-trained foundation model is comparable to or even better than some traditional recommendation models based on supervised training. The IDGenRec framework addresses the challenge of encoding items as language tokens in generative recommendation systems, enabling the use of LLMs to better align with recommendation needs. The framework's ability to generate meaningful textual IDs and its performance in both standard and zero-shot settings highlight its potential as a foundation model for future recommendation systems.IDGenRec: LLM-RecSys Alignment with Textual ID Learning This paper proposes IDGenRec, a novel framework for aligning large language models (LLMs) with recommendation systems by learning textual IDs for items. The framework generates unique, concise, and semantically rich textual IDs using human language tokens, enabling seamless integration of personalized recommendations into natural language generation. The ID generator is trained alongside the LLM-based recommender, allowing for collaborative learning and improved recommendation performance. The ID generator produces item IDs based on item metadata, ensuring that the IDs are short, unique, and meaningful for recommendation purposes. A diverse ID generation algorithm is employed to ensure uniqueness, and the generated IDs are integrated into the text-to-text generation process. The base recommender then generates the target item ID based on the generated IDs and user history. The framework is evaluated on four standard sequential recommendation datasets and shows significant improvements over existing baselines. Additionally, the framework is tested in a zero-shot setting on six unseen datasets, demonstrating strong performance and the potential of IDGenRec as a foundation model for generative recommendation. The results show that the zero-shot performance of the pre-trained foundation model is comparable to or even better than some traditional recommendation models based on supervised training. The IDGenRec framework addresses the challenge of encoding items as language tokens in generative recommendation systems, enabling the use of LLMs to better align with recommendation needs. The framework's ability to generate meaningful textual IDs and its performance in both standard and zero-shot settings highlight its potential as a foundation model for future recommendation systems.