Envision Recommendation on LLM-based Agent Platform
Abstract
Large Language Model (LLM)-based platforms like GPTs form a new kind of Agent-oriented information system, pushing us to scrutinize the infrastructure of information systems to support Agent-level information processing and accommodate properties of LLM-based Agents such as interactivity. In this work, we argue for the prospects for the recommender system for LLM-based Agent platforms and propose a novel recommendation paradigm named the Rec4Agentverse, which consists of Item Agents and the Agent Recommender. In this paradigm, we highlight the collaboration between Item Agents and the Agent Recommender to facilitate personalized information service, and enrich information exchange beyond the conventional feedback loop between the user and the recommender. Furthermore, we envision the development of the Rec4Agentverse and conceptualize it into three stages with enriching interaction and information exchange among the user, the Agent Recommender, and Item Agents. A preliminary study with several cases for instantiating the Rec4Agentverse demonstrates its considerable potential for application, followed by a thorough discussion of possible issues and potential research directions.
Citation:
@article{zhang2025envision,
title={Envision Recommendation on LLM-based Agent Platform},
author={Zhang, Jizhi and Bao, Keqin and Wang, Wenjie and Zhang, Yang and Shi, Wentao and Xu, Wanhong and Feng, Fuli and Chua, Tat-Seng},
journal={Communications of the ACM},
year={2025}
}
Jizhi Zhang
张及之
Keqin Bao
鲍克勤
Wenjie Wang
王文杰 教授
Wentao Shi
石文焘
