# Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach

**Authors:** Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, Ji-Rong Wen  
**arXiv:** [2305.07001](https://arxiv.org/abs/2305.07001)  
**Date:** 2023-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.07001v1/

## Key points

- Recommender systems have gained popularity and attention in research and industry.
- A new approach to recommendation models has emerged, inspired by large language models (LLMs).
- LLMs can understand and execute natural language instructions for recommendation tasks.
- The researchers chose to instruction tune the open-source LLM called 3B Flan-T5-XL for recommender systems.
- They designed a general instruction format and created 39 instruction templates to generate a dataset of user-personalized instructions.
- Extensive experiments using real-world datasets showed that their approach outperformed competitive baselines, including GPT-3.5.
- This study demonstrates the potential for developing user-friendly recommender systems through natural language instructions.
- The researchers optimized 3B Flan-T5-XL through instruction tuning using a sequence-to-sequence framework.
- Inference quality was evaluated by comparing it to text davinci 003 - a teacher LLM - and considering agreement percentages related to user's information provided by the teacher LLM.
- This study was presented at Conference acronym 'XX' in August 2022 in Woodstock NY.

## Summary

In recent years, recommender systems have become increasingly popular and have been the focus of numerous research studies and industry developments. These systems aim to provide effective recommendations by learning user preferences from historical behavior data and establishing user-item matching relationships. However, a new approach to developing recommendation models has emerged, inspired by the advancements in large language models (LLMs). This approach views recommendation as instruction following by LLMs. It recognizes that users can express their preferences or needs through natural language descriptions, which LLMs can understand and execute to fulfill the recommendation task. To better adapt LLMs to recommender systems, the researchers in this study chose to instruction tune an open-source LLM called 3B Flan-T5-XL. To facilitate instruction tuning, the researchers first designed a general instruction format that describes a user's preference, intention, task form, and context in natural language. They then manually created 39 instruction templates and automatically generated a large dataset of user-personalized instructions (252K instructions) with varying types of preferences and intentions. To evaluate the effectiveness of their approach, the researchers instantiated the instruction templates into several well-known recommendation tasks and conducted extensive experiments using real-world datasets. The results showed that their approach outperformed several competitive baselines, including GPT-3.5. The findings of this study demonstrate the potential for developing more user-friendly recommender systems that allow users to communicate freely with the system and receive more accurate recommendations through natural language instructions. To optimize LLMs for recommender systems through instruction tuning, the researchers introduced 3B Flan-T5-XL as its backbone model due to its fine-tuning based on T5 with a large amount of instruction data. The training process involved optimizing negative log likelihoods of target outputs based on different types of instructions unified into a sequence-to-sequence framework while inference was evaluated for quality by comparing them to text davinci 003 - a teacher LLM - considering agreement percentages related to user's related information provided by teacher LLM as well as reflecting user's preference or intention in instructions generated from it. Overall, this study provides insights into instruction tuning for recommender systems using LLMs and highlights its potential for developing more effective and user friendly recommendation models presented at Conference acronym 'XX' in August 2022 in Woodstock NY.