**USER-LLM: Efficient LLM Contextualization with User Embeddings** **Abstract:** Large language models (LLMs) have revolutionized natural language processing, but incorporating complex and noisy user interaction data remains challenging. To address this, the authors propose USER-LLM, a framework that leverages user embeddings to contextualize LLMs. These embeddings, distilled from diverse user interactions using self-supervised pre-training, capture latent user preferences and their evolution over time. USER-LLM integrates these user embeddings with LLMs through cross-attention and soft-prompting, enabling dynamic adaptation to user context. Experiments on MovieLens, Amazon Review, and Google Local Review datasets demonstrate significant performance gains across various tasks, particularly in long sequence tasks and those requiring deep user understanding. The approach outperforms text-prompt-based contextualization while being computationally efficient. Perceiver layers are incorporated to streamline the integration between user encoders and LLMs, further reducing computational demands. **Introduction:** User interactions provide rich behavioral data, but leveraging this data with LLMs is complex due to the complexity, sparsity, and multimodality of user interactions. USER-LLM addresses these challenges by generating high-quality user embeddings from diverse modalities and integrating them into LLMs. The framework consists of two phases: generating user embeddings and contextualizing LLMs with these embeddings. The authors evaluate USER-LLM on various tasks and datasets, demonstrating its effectiveness in personalization and computational efficiency. **USER-LLM:** USER-LLM uses a Transformer-based encoder to generate user embeddings from user interaction data. These embeddings are then integrated into the LLM through cross-attention, enabling dynamic context injection. The framework offers flexible training strategies, including full finetuning, encoders-only finetuning, LoRA-based tuning, and projection layers-only finetuning. Experiments show that USER-LLM outperforms baselines in tasks requiring deep user understanding and long context inputs, while maintaining computational efficiency. **Experiments:** USER-LLM is evaluated on MovieLens, Amazon Review, and Google Local Review datasets. Results show significant performance improvements over non-LLM baselines and text-prompt-based LLM personalization techniques. USER-LLM demonstrates strong generalization capabilities in tasks requiring deep user understanding and efficient inference. **Conclusion:** USER-LLM is a versatile framework that leverages user embeddings to contextualize LLMs, enhancing their ability to understand and adapt to user contexts. The approach is computationally efficient and suitable for real-world applications. Future work could focus on optimizing user embedding generation, improving context understanding, and expanding the framework's adaptability.**USER-LLM: Efficient LLM Contextualization with User Embeddings** **Abstract:** Large language models (LLMs) have revolutionized natural language processing, but incorporating complex and noisy user interaction data remains challenging. To address this, the authors propose USER-LLM, a framework that leverages user embeddings to contextualize LLMs. These embeddings, distilled from diverse user interactions using self-supervised pre-training, capture latent user preferences and their evolution over time. USER-LLM integrates these user embeddings with LLMs through cross-attention and soft-prompting, enabling dynamic adaptation to user context. Experiments on MovieLens, Amazon Review, and Google Local Review datasets demonstrate significant performance gains across various tasks, particularly in long sequence tasks and those requiring deep user understanding. The approach outperforms text-prompt-based contextualization while being computationally efficient. Perceiver layers are incorporated to streamline the integration between user encoders and LLMs, further reducing computational demands. **Introduction:** User interactions provide rich behavioral data, but leveraging this data with LLMs is complex due to the complexity, sparsity, and multimodality of user interactions. USER-LLM addresses these challenges by generating high-quality user embeddings from diverse modalities and integrating them into LLMs. The framework consists of two phases: generating user embeddings and contextualizing LLMs with these embeddings. The authors evaluate USER-LLM on various tasks and datasets, demonstrating its effectiveness in personalization and computational efficiency. **USER-LLM:** USER-LLM uses a Transformer-based encoder to generate user embeddings from user interaction data. These embeddings are then integrated into the LLM through cross-attention, enabling dynamic context injection. The framework offers flexible training strategies, including full finetuning, encoders-only finetuning, LoRA-based tuning, and projection layers-only finetuning. Experiments show that USER-LLM outperforms baselines in tasks requiring deep user understanding and long context inputs, while maintaining computational efficiency. **Experiments:** USER-LLM is evaluated on MovieLens, Amazon Review, and Google Local Review datasets. Results show significant performance improvements over non-LLM baselines and text-prompt-based LLM personalization techniques. USER-LLM demonstrates strong generalization capabilities in tasks requiring deep user understanding and efficient inference. **Conclusion:** USER-LLM is a versatile framework that leverages user embeddings to contextualize LLMs, enhancing their ability to understand and adapt to user contexts. The approach is computationally efficient and suitable for real-world applications. Future work could focus on optimizing user embedding generation, improving context understanding, and expanding the framework's adaptability.