The paper addresses the challenge of aligning large language models (LLMs) with recommendation systems, particularly in tasks such as retrieval, ranking, and rating prediction. The authors identify a gap between LLMs' strengths in natural language reasoning and their inability to model complex user-item interactions, which are crucial for effective recommendations. To bridge this gap, they propose a novel approach that involves fine-tuning LLMs with auxiliary-task data samples and recommendation-task data samples. **Auxiliary-Task Data Samples:** - **Masked Item Modeling (MIM):** Encodes item correlations by predicting masked items in user purchase sequences. - **Masked Language Modeling (MLM):** Models fine-grained item correlations by masking consecutive spans of tokens. - **Bayesian Personalized Ranking (BPR):** Learns user preferences by contrasting positive and negative items. **Recommendation-Task Data Samples:** - **Reduces complexity by eliminating user IDs and using item titles.** - **Includes raw user purchase sequences to enhance sequence modeling.** **Experiments:** - **Datasets:** Amazon Toys & Games, Beauty, and Sports & Outdoors. - **Tasks:** Retrieval, ranking, and rating prediction. - **Results:** The proposed method significantly outperforms conventional and LLM-based baselines, including the current state-of-the-art (SOTA) model, TIGER, in retrieval tasks. It also shows promising performance in ranking and rating prediction tasks, though with some limitations in Sports & Outdoors due to the nature of the data. **Contributions:** - Introduces a novel method to align LLMs with new recommendation domains. - Proposes more informative recommendation-task data samples compared to existing studies. - Demonstrates the effectiveness of the proposed method through extensive experiments. **Limitations:** - Computational resource consumption and extended training times due to the large parameter size of LLMs. The paper highlights the potential of LLMs in recommendation systems and suggests future directions for further research, including addressing novel recommendation scenarios and leveraging the diverse capabilities of LLM backbones.The paper addresses the challenge of aligning large language models (LLMs) with recommendation systems, particularly in tasks such as retrieval, ranking, and rating prediction. The authors identify a gap between LLMs' strengths in natural language reasoning and their inability to model complex user-item interactions, which are crucial for effective recommendations. To bridge this gap, they propose a novel approach that involves fine-tuning LLMs with auxiliary-task data samples and recommendation-task data samples. **Auxiliary-Task Data Samples:** - **Masked Item Modeling (MIM):** Encodes item correlations by predicting masked items in user purchase sequences. - **Masked Language Modeling (MLM):** Models fine-grained item correlations by masking consecutive spans of tokens. - **Bayesian Personalized Ranking (BPR):** Learns user preferences by contrasting positive and negative items. **Recommendation-Task Data Samples:** - **Reduces complexity by eliminating user IDs and using item titles.** - **Includes raw user purchase sequences to enhance sequence modeling.** **Experiments:** - **Datasets:** Amazon Toys & Games, Beauty, and Sports & Outdoors. - **Tasks:** Retrieval, ranking, and rating prediction. - **Results:** The proposed method significantly outperforms conventional and LLM-based baselines, including the current state-of-the-art (SOTA) model, TIGER, in retrieval tasks. It also shows promising performance in ranking and rating prediction tasks, though with some limitations in Sports & Outdoors due to the nature of the data. **Contributions:** - Introduces a novel method to align LLMs with new recommendation domains. - Proposes more informative recommendation-task data samples compared to existing studies. - Demonstrates the effectiveness of the proposed method through extensive experiments. **Limitations:** - Computational resource consumption and extended training times due to the large parameter size of LLMs. The paper highlights the potential of LLMs in recommendation systems and suggests future directions for further research, including addressing novel recommendation scenarios and leveraging the diverse capabilities of LLM backbones.