A Multi-Agent Conversational Recommender System

TL;DR

The paper addresses the challenge of controlling dialogue flow and leveraging user feedback in conversational recommender systems (CRS) by proposing MACRS, an LLM-only framework that combines a multi-agent act planning module with three responder agents and one planner, and a user feedback-aware reflection mechanism operating at information-level and strategy-level. The system dynamically plans dialogue acts, generates multiple candidate responses, and selects the most appropriate one, while reflection updates user profiles and dialogue strategies based on feedback across turns. Ablation studies and experiments on MovieLens show MACRS improves recommendation accuracy and efficiency of preference elicitation compared to strong LLM baselines and traditional CRS approaches. The work highlights the practical potential of integrated, feedback-driven, multi-agent control for enhancing user experience in interactive recommendation settings.

Abstract

Due to strong capabilities in conducting fluent, multi-turn conversations with users, Large Language Models (LLMs) have the potential to further improve the performance of Conversational Recommender System (CRS). Unlike the aimless chit-chat that LLM excels at, CRS has a clear target. So it is imperative to control the dialogue flow in the LLM to successfully recommend appropriate items to the users. Furthermore, user feedback in CRS can assist the system in better modeling user preferences, which has been ignored by existing studies. However, simply prompting LLM to conduct conversational recommendation cannot address the above two key challenges. In this paper, we propose Multi-Agent Conversational Recommender System (MACRS) which contains two essential modules. First, we design a multi-agent act planning framework, which can control the dialogue flow based on four LLM-based agents. This cooperative multi-agent framework will generate various candidate responses based on different dialogue acts and then choose the most appropriate response as the system response, which can help MACRS plan suitable dialogue acts. Second, we propose a user feedback-aware reflection mechanism which leverages user feedback to reason errors made in previous turns to adjust the dialogue act planning, and higher-level user information from implicit semantics. We conduct extensive experiments based on user simulator to demonstrate the effectiveness of MACRS in recommendation and user preferences collection. Experimental results illustrate that MACRS demonstrates an improvement in user interaction experience compared to directly using LLMs.

Figures (7)

• Figure 1: Example of modeling a Conversational Recommendation System (CRS) using a multi-agent framework. The right side shows a dialogue between CRS and user, while the left side depicts the internal framework of MACRS. The responder agent and planner agent collaboratively generate appropriate responses, while the reflection mechanism provides feedback and refined guidance to these agents, optimizing their responses to better satisfy user needs.
• Figure 2: The overview of MACRS, which contains two modules: multi-agent act planning and user feedback-aware reflection. The multi-agent act planning module generates the response according to the user feedback. The user feedback-aware reflection module summarizes the high-level user profile and dialogue strategy suggestions which helps the responder agents to optimize their response.
• Figure 3: Overview of user feedback-aware reflection mechanism.
• Figure 4: The ratio of different dialogue acts chosen in each turn of dialogue. The diversity in dialogue act selection of our proposed MACRS is remarkably enhanced compared to the corresponding backbone model, and it enhances the user experience.
• Figure 5: Cumulative number of successful samples for each dialogue turn.