Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PERSOMA: PERsonalized SOft ProMpt Adapter Architecture for Personalized Language Prompting

Liam Hebert, Krishna Sayana, Ambarish Jash, Alexandros Karatzoglou, Sukhdeep Sodhi, Sumanth Doddapaneni, Yanli Cai, Dima Kuzmin

TL;DR

PERSOMA addresses the challenge of personalizing large language models by compressing extensive user histories into expressive soft prompt embeddings using a history encoder (SentenceT5) and a learnable soft-prompt adapter, followed by a resampling step to a compact token set. The prompts P_u are combined with fixed task prompts T and fed into a frozen LLM (PaLM 2) to generate personalized outputs, enabling substantial personalization without full model finetuning. On the MovieLens Genre Prediction task, PERSOMA outperforms embedding-based baselines and matches end-to-end text prompting while drastically reducing input and compute; LoRA-based parameter-efficient training approaches further close the gap to full finetuning. The results highlight the effectiveness of history resampling and expressive prompt embeddings for scalable personalization, with future work exploring larger histories, sparse retrieval, and diverse LLM backbones.

Abstract

Understanding the nuances of a user's extensive interaction history is key to building accurate and personalized natural language systems that can adapt to evolving user preferences. To address this, we introduce PERSOMA, Personalized Soft Prompt Adapter architecture. Unlike previous personalized prompting methods for large language models, PERSOMA offers a novel approach to efficiently capture user history. It achieves this by resampling and compressing interactions as free form text into expressive soft prompt embeddings, building upon recent research utilizing embedding representations as input for LLMs. We rigorously validate our approach by evaluating various adapter architectures, first-stage sampling strategies, parameter-efficient tuning techniques like LoRA, and other personalization methods. Our results demonstrate PERSOMA's superior ability to handle large and complex user histories compared to existing embedding-based and text-prompt-based techniques.

PERSOMA: PERsonalized SOft ProMpt Adapter Architecture for Personalized Language Prompting

TL;DR

PERSOMA addresses the challenge of personalizing large language models by compressing extensive user histories into expressive soft prompt embeddings using a history encoder (SentenceT5) and a learnable soft-prompt adapter, followed by a resampling step to a compact token set. The prompts P_u are combined with fixed task prompts T and fed into a frozen LLM (PaLM 2) to generate personalized outputs, enabling substantial personalization without full model finetuning. On the MovieLens Genre Prediction task, PERSOMA outperforms embedding-based baselines and matches end-to-end text prompting while drastically reducing input and compute; LoRA-based parameter-efficient training approaches further close the gap to full finetuning. The results highlight the effectiveness of history resampling and expressive prompt embeddings for scalable personalization, with future work exploring larger histories, sparse retrieval, and diverse LLM backbones.

Abstract

Understanding the nuances of a user's extensive interaction history is key to building accurate and personalized natural language systems that can adapt to evolving user preferences. To address this, we introduce PERSOMA, Personalized Soft Prompt Adapter architecture. Unlike previous personalized prompting methods for large language models, PERSOMA offers a novel approach to efficiently capture user history. It achieves this by resampling and compressing interactions as free form text into expressive soft prompt embeddings, building upon recent research utilizing embedding representations as input for LLMs. We rigorously validate our approach by evaluating various adapter architectures, first-stage sampling strategies, parameter-efficient tuning techniques like LoRA, and other personalization methods. Our results demonstrate PERSOMA's superior ability to handle large and complex user histories compared to existing embedding-based and text-prompt-based techniques.
Paper Structure (12 sections, 2 equations, 1 figure, 6 tables)

This paper contains 12 sections, 2 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. PERSOMA Architecture
  5. MovieLens Personalized Genre Prediction Task
  6. Results
  7. Effect of PERSOMA Soft Prompt Adapter Architecture
  8. Parameter Efficient Training with LoRA
  9. Effect of History Sampling Strategies
  10. Conclusions & Future Work
  11. Label distribution
  12. Effect of Item Input Format

Figures (1)

  • Figure 1: Overview of the PERSOMA architecture. Sentence T5 encodes natural language descriptions of the user's history and is then jointly resampled by a Perceiver adapter to create in-context user soft prompts. These soft prompts are then concatenated with task embeddings to prompt a PaLM 2 language model towards personalized output.