# LaMP: When Large Language Models Meet Personalization

**Authors:** Alireza Salemi, Sheshera Mysore, Michael Bendersky, Hamed Zamani  
**arXiv:** [2304.11406](https://arxiv.org/abs/2304.11406)  
**Date:** 2023-04-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.11406v3/

## Key points

- The field of NLP is rapidly evolving, highlighting the need for personalization in language models.
- Users today seek tailored experiences aligned with their unique needs and preferences, emphasizing the importance of personalized responses from LLMs.
- There is a gap in developing and evaluating LLMs for producing personalized outputs across various research communities.
- The LaMP benchmark is introduced to bridge this gap, offering a comprehensive evaluation framework with diverse language tasks and multiple entries for each user profile.
- LaMP consists of seven personalized tasks including text classification tasks and text generation tasks.
- Two retrieval augmentation approaches are proposed to personalize language model outputs by retrieving personal items from each user profile.
- Extensive experiments on LaMP demonstrate the effectiveness of the proposed retrieval augmentation approach for both zero-shot and fine-tuned language models.
- LaMP fills a crucial gap in existing NLP benchmarks by focusing on personalization as a key factor in shaping the future of NLP systems.

## Summary

The field of  is rapidly evolving and the need for personalization in  has become increasingly apparent. Users today seek tailored experiences that align with their unique needs and preferences, making personalized responses from LLMs crucial. However, despite widespread interest in personalization across various research communities such as information retrieval and human-computer interaction, there remains a gap in developing and evaluating LLMs for producing personalized outputs. To bridge this gap, this paper introduces the , a novel benchmark specifically designed to train and evaluate language models for personalized outputs. It offers a comprehensive evaluation framework with diverse language tasks and multiple entries for each user profile. The benchmark consists of seven personalized tasks including text classification tasks like , , and , as well as four text generation tasks. In addition to presenting these tasks, the paper proposes two retrieval augmentation approaches that retrieve personal items from each user profile to personalize language model outputs. Various retrieval models are studied, including term matching, semantic matching, and time-aware methods. Extensive experiments on LaMP demonstrate the effectiveness of the proposed retrieval augmentation approach for both zero-shot and fine-tuned language models. The introduction of LaMP fills a crucial gap in existing NLP benchmarks by focusing on personalization as a key factor in shaping the future of NLP systems. While well-known benchmarks like GLUE and SuperGLUE have driven progress in various NLP tasks, they often adopt a "one-size-fits-all" approach that limits research on personalization. In contrast,< kd>LaMP provides a platform for developing models that can adapt to specific user needs across personalized text classification and generation tasks</kd>. Overall, this paper underscores the importance of personalization in NLP systems and takes an important step towards advancing research on personalized responses from large language models through the introduction of the LaMP benchmark.