Beyond Retrieval: Learning Compact User Representations for Scalable LLM Personalization

The article presents TAP-PER (Temporal Attentive Prefix for PERsonalization), a novel framework for personalizing large language models by learning compact user representations through prefix-based embeddings. TAP-PER significantly reduces the parameter requirements, utilizing 130 times fewer per-user parameters compared to OPPU and approximately half the total parameter footprint of PER-PCS at a scale of 1,000 users. This approach enhances scalability and efficiency in LLM personalization, addressing the limitations of existing methods that rely on retrieval quality or extensive user-specific parameter storage.