Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Embedding-to-Prefix: Parameter-Efficient Personalization for Pre-Trained Large Language Models

Bernd Huber, Ghazal Fazelnia, Andreas Damianou, Sebastian Peleato, Max Lefarov, Praveen Ravichandran, Marco De Nadai, Mounia Lalmas-Roellke, Paul N. Bennett

TL;DR

Embedding-to-Prefix (E2P) addresses the need for efficient personalization of large language models by mapping pre-computed dense user embeddings into a single soft prefix token, generated through a lightweight projection, while keeping the LLM frozen. This parameter-efficient approach yields substantial improvements across dialogue, news headline generation, and production-scale music and podcast personalization, with minimal latency and training requirements. The method demonstrates strong generalization, robustness to embedding drift, and complementary benefits when combined with textual prompts, suggesting practical viability for real-world personalization at scale. Overall, E2P provides a scalable pathway to personalized generation without full model fine-tuning, broadening the applicability of LLMs in individualized user experiences.

Abstract

Large language models (LLMs) excel at generating contextually relevant content. However, tailoring these outputs to individual users for effective personalization is a significant challenge. While rich user-specific information often exists as pre-existing user representations, such as embeddings learned from preferences or behaviors, current methods to leverage these for LLM personalization typically require costly fine-tuning or token-heavy prompting. We propose Embedding-to-Prefix (E2P), a parameter-efficient method that injects pre-computed context embeddings into an LLM's hidden representation space through a learned projection to a single soft token prefix. This enables effective personalization while keeping the backbone model frozen and avoiding expensive adaptation techniques. We evaluate E2P across two public datasets and in a production setting: dialogue personalization on Persona-Chat, contextual headline generation on PENS, and large-scale personalization for music and podcast consumption. Results show that E2P preserves contextual signals and achieves strong performance with minimal computational overhead, offering a scalable, efficient solution for contextualizing generative AI systems.

Embedding-to-Prefix: Parameter-Efficient Personalization for Pre-Trained Large Language Models

TL;DR

Embedding-to-Prefix (E2P) addresses the need for efficient personalization of large language models by mapping pre-computed dense user embeddings into a single soft prefix token, generated through a lightweight projection, while keeping the LLM frozen. This parameter-efficient approach yields substantial improvements across dialogue, news headline generation, and production-scale music and podcast personalization, with minimal latency and training requirements. The method demonstrates strong generalization, robustness to embedding drift, and complementary benefits when combined with textual prompts, suggesting practical viability for real-world personalization at scale. Overall, E2P provides a scalable pathway to personalized generation without full model fine-tuning, broadening the applicability of LLMs in individualized user experiences.

Abstract

Large language models (LLMs) excel at generating contextually relevant content. However, tailoring these outputs to individual users for effective personalization is a significant challenge. While rich user-specific information often exists as pre-existing user representations, such as embeddings learned from preferences or behaviors, current methods to leverage these for LLM personalization typically require costly fine-tuning or token-heavy prompting. We propose Embedding-to-Prefix (E2P), a parameter-efficient method that injects pre-computed context embeddings into an LLM's hidden representation space through a learned projection to a single soft token prefix. This enables effective personalization while keeping the backbone model frozen and avoiding expensive adaptation techniques. We evaluate E2P across two public datasets and in a production setting: dialogue personalization on Persona-Chat, contextual headline generation on PENS, and large-scale personalization for music and podcast consumption. Results show that E2P preserves contextual signals and achieves strong performance with minimal computational overhead, offering a scalable, efficient solution for contextualizing generative AI systems.

Paper Structure

This paper contains 27 sections, 5 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Simplified Prefix Token
  5. Projection Module
  6. Soft Token Insertion
  7. Training Objectives
  8. Text Generation Tasks
  9. Recommendation Tasks
  10. Experimental Setup
  11. Datasets
  12. Models and Training
  13. Evaluation Metrics
  14. Results
  15. Analysis of Results
  16. ...and 12 more sections

Figures (2)

  • Figure 1: Overview of Embedding-to-Prefix (E2P). A projection module maps a user embedding $c$ to a soft prefix token $p$ that is prepended to the LLM input embedding sequence. The soft prefix influences the entire generation process without modifying the frozen LLM weights.
  • Figure 2: Visualization of user embedding properties in the LLM's hidden representation space: (a) t-SNE visualization of user embeddings colored by music genre preference clusters, (b) Corresponding visualization of the same users in prefix adapter space, demonstrating preservation of preference clusters with enhanced separation between distinct user groups. This plot demonstrates that the learned projection preserves meaningful user preference structures (e.g., genre clusters) when mapping from the user embedding space to the prefix adapter space, supporting the potential for generalization based on user similarity.