# A Large Language Model Enhanced Conversational Recommender System

**Authors:** Yue Feng, Shuchang Liu, Zhenghai Xue, Qingpeng Cai, Lantao Hu, Peng Jiang, Kun Gai, Fei Sun  
**arXiv:** [2308.06212](https://arxiv.org/abs/2308.06212)  
**Date:** 2023-08-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.06212v1/

## Key points

- Conversational Recommender Systems (CRSs) recommend high-quality items to users through a dialogue interface
- Large Language Models (LLMs) have remarkable reasoning and generation capabilities, enhancing CRSs
- LLM-based CRS called LLMCRS addresses challenges by leveraging LLM reasoning ability and expert models
- LLMCRS system stages: sub-task detection, model matching, sub-task execution, response generation
- Fine-tuning LLM using reinforcement learning from CRS performance feedback (RLPF) improves recommendation accuracy
- Experimental results show that LLMCRS with RLPF outperforms existing methods
- Evaluation measures include HIT@k, MRR@k, NDCG@k for recommendation performance assessment; BLEU for response similarity evaluation; Distinct for response diversity measurement
- Integration of LLM capabilities in conversational recommender systems improves recommendation accuracy and enhances user interactions

## Summary

, , , , 


Conversational Recommender Systems (CRSs) are designed to recommend high-quality items to users through a dialogue interface, involving tasks such as user preference elicitation, recommendation, explanation, and item information search. Developing effective CRSs poses challenges in managing sub-tasks, solving them effectively, and generating responses that engage users. Large Language Models (LLMs) have shown remarkable reasoning and generation capabilities, offering new opportunities for enhancing CRSs. In this study, a novel LLM-based CRS called LLMCRS is proposed to address these challenges. are designed to recommend high-quality items to users through a dialogue interface. LLMCRS leverages the reasoning ability of LLM for sub-task management and collaborates with expert models for enhanced performance in solving different sub-tasks. The system is divided into four stages: sub-task detection, model matching, sub-task execution, and response generation. It incorporates schema-based instruction, demonstration-based instruction, dynamic sub-task and model matching, and summary-based generation to guide LLM in generating desired results. To adapt LLM to conversational recommendations, the system is fine-tuned using reinforcement learning from CRS performance feedback (RLPF). Experimental results on benchmark datasets demonstrate that LLMCRS with RLPF outperforms existing methods. have shown remarkable reasoning and generation capabilities. A case study comparing LLMCRS with TG-ReDial on GoRecDial datasets illustrates the system's ability to provide accurate recommendations and generate informative responses based on user preferences. Evaluation measures include ranking-based metrics like HIT@k, MRR@k, NDCG@k for recommendation performance assessment; relevance-based metrics like BLEU for response similarity evaluation; and diversity-based metrics like Distinct for response diversity measurement. Implementation details ensure fair comparisons using the CRSLab toolkit. showcases the potential of integrating LLM capabilities into conversational recommender systems to improve recommendation accuracy and enhance user interactions through more informative responses. Overall, LLMCRS demonstrates the potential of  in improving conversational recommender systems. It offers a novel approach to managing sub-tasks, solving them effectively, and generating engaging responses that cater to user preferences. By leveraging the capabilities of , LLMCRS presents new opportunities for enhancing CRSs and providing accurate recommendations based on user preferences. Through its innovative design and implementation, LLMCRS showcases the importance of considering user interactions in developing effective conversational recommender systems.