Satisfying Complex User Needs: M^3 Agent for Conversational Multi-Item Recommendation

Dazhou Yu, Yanhui Guo, Jingjing Meng, Alireza Bagheri Garakani, Liang Zhao, Yan Gao

Keywords: agent, LLM, shopping, ecommerce, multi-item, recommendation, search

TL;DR: a multi-objective optimization framework that bridges natural-language requests and catalog-grounded, constraint-checked multi-item recommendations

Abstract: Fulfilling complex, multi-item, multi-constraint user requests in conversational commerce is a critical and largely unsolved research problem. Existing paradigms, from traditional recommenders to modern LLM-based agents, fail due to a crisis of grounding and an inability to handle trade-off complexity. To address these failures, we introduce M^3Agent, an agentic framework that bridges natural language and grounded, optimal recommendations. We are the first to reformulate this task as a unified multi-objective optimization problem, where the agent's plan is the provably optimal solution to a holistic objective. M^3Agent's cognitive architecture employs two core mechanisms: an Split-Prune Constraint Tree for grounded, interactive memory, and a Pareto-complete search to find all optimal trade-offs between competing quality objectives. Experiments on two large-scale, real-world datasets show consistent and significant gains over strong baselines, demonstrating that M^3Agent effectively translates free-text multi-item requests into catalog-valid, requirement-satisfying recommendations.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 4912