MCTA4R: Enhancing User Preference Alignment with Meta-CoT and Search Algorithms
Ruilong Huang, Bohan Li, Xinzhe Zhao, Mengfei Xu, Wenlong Wu, Qi Zhu · 2025
Personalized recommendation systems often struggle with accurately predicting long-term user preferences, particularly in dynamic and sparse environments. Aligning user preferences is a key issue in LLM-driven recommendation systems with generative agents. To address this, we propose the Meta-CoT-Agent4Rec (MCTA4R) framework, which integrates the Meta-CoT reasoning process with$A^{*}$and Monte Carlo Tree Search (MCTS) algorithms. The$A^{*}$search algorithm leverages a heuristic strategy to guide the model towards optimal recommendations by expanding nodes based on cost evaluations, ensuring relevance in structured data environments. Meanwhile, the MCTS utilizes random simulations and backpropagation to explore diverse recommendation paths, adapting to dynamic and sparse data scenarios. The framework was evaluated using the MovieLens-1M, MovieLens-10M, Amazon-Book, and Steam datasets with different interaction ratios. Experimental results demonstrate that both$\mathrm{A}^{*}$Meta-CoT and MCTS Meta-CoT significantly outperform traditional few-shot learning methods in terms of accuracy, recall and F1 score.