Large Language Model Agents for Recommender Systems: Bridging Behavior and Semantics with Long-Short Term Interest Modeling

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A recommender system extracts user preferences from past interactions to suggest items, widely used in platforms like user-generated content, online shopping, and urban services. These systems aim to provide accurate recommendations, reduce user interaction burden, and enhance user experience while improving socio-economic benefits. Current technologies include content-based, behavior-based, and hybrid recommendations. Behavior-based technologies, like collaborative filtering, are mature due to extensive user–item interaction data but focus on behavior over intrinsic features, lacking comprehensive understanding. Advancements in natural language processing (NLP), text vectorization, and large language models (LLMs) have made content-based recommendations based on semantic understanding more prominent. This study presents an exploratory LLM-agent-based hybrid recommender system framework, building upon a multi-armed bandit (MAB) approach, which enables semantic understanding of item content and integrates both short-term and long-term user interaction behaviors for recommendation. Specifically, it optimizes collaborative filtering models to capture long-term and short-term interests, and uses an agentic LLM memory and reasoning component to identify preference shifts and schedule appropriate recommendation modules. It further vectorizes content using NLP and LLMs to generate retrieval terms that enhance candidate recall, and aligns LLM-optimized content recommendations with user behavior to integrate both aspects. This hybrid system achieved a recall rate improvement of up to 26.63% in behavior recommendation and, in ranking-oriented (diversity-focused) recommendation tasks on the MovieLens dataset, raised NDCG@100 from 0.1428 to 0.3541, a relative improvement of approximately 148% (i.e., about 2.48× the baseline value), albeit against a deliberately simple multi-armed bandit baseline and with a corresponding decrease in Recall@k. However, this work represents an initial exploration in this emerging research direction, and more comprehensive comparisons with state-of-the-art methods are needed in future work to fully assess the framework’s effectiveness.