# Teach LLMs to Personalize -- An Approach inspired by Writing Education

**Authors:** Cheng Li, Mingyang Zhang, Qiaozhu Mei, Yaqing Wang, Spurthi Amba Hombaiah, Yi Liang, Michael Bendersky  
**arXiv:** [2308.07968](https://arxiv.org/abs/2308.07968)  
**Date:** 2023-08-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.07968v1/

## Key points

- Personalized text generation using large language models (LLMs) is an emerging research area
- Existing studies in this field often focus on specific domains and use bespoke features or models
- The proposed approach leverages LLMs and draws inspiration from writing education practices
- The authors develop a multistage and multitask framework for teaching LLMs to generate personalized text
- The stages of the approach include retrieval, ranking, summarization, synthesis, and generation
- A multitask setting is introduced to enhance the model's generation ability based on the correlation between reading proficiency and writing ability in education
- The approach is evaluated on three public datasets and shows significant improvements over various baselines.

## Summary

The paper titled "Teach LLMs to Personalize -- An Approach inspired by Writing Education" proposes a general approach for personalized text generation using large language models (LLMs). The authors highlight that personalized text generation is an emerging research area that has gained significant attention in recent years. Most existing studies in this field focus on specific domains and rely on bespoke features or models. In contrast, the proposed approach leverages LLMs and draws inspiration from writing education practices. To teach LLMs for personalized generation, the authors develop a multistage and multitask framework. They draw an analogy with the task of writing from sources in education, which involves multiple steps such as finding, evaluating, summarizing, synthesizing, and integrating information. Similarly, their approach to personalized text generation consists of several stages: retrieval, ranking, summarization, synthesis and generation. Additionally, the authors introduce a multitask setting to further enhance the model's generation ability. This idea is motivated by the observation that reading proficiency and writing ability are often correlated in education. By evaluating their approach on three public datasets covering different domains, the authors demonstrate significant improvements over various baselines. Overall, this paper presents a novel approach for personalized text generation using LLMs which draws inspiration from writing education practices. The proposed multistage and multitask framework shows promising results in improving the model's ability to generate personalized text across different domains.