Large language model-based agents rely heavily on memory to achieve personalization by storing and utilizing user information. Previous studies have mainly focused on preference alignment and simple question-answering tasks, which do not fully capture the complexity of real-world scenarios requiring multi-hop reasoning over extensive personalized data. To address this gap, the authors define a multi-hop personalized reasoning task and construct a corresponding dataset along with a unified evaluation framework. They implement and compare various explicit and implicit memory methods, analyzing their strengths and weaknesses in handling complex reasoning tasks. Recognizing the limitations of both paradigms, the authors propose a hybrid approach called HybridMem that combines explicit and implicit memory mechanisms to improve performance. Extensive experiments demonstrate the effectiveness of HybridMem in overcoming challenges posed by multi-hop reasoning over personalized information. This work highlights the importance of advanced memory mechanisms for personalization in language model agents and provides a benchmark for future research. The authors have made their dataset, code, and evaluation framework publicly available to facilitate further exploration and development in this area. This contribution is significant as it pushes the boundaries of personalized AI by addressing complex reasoning tasks that better reflect real-world user interactions.
👉 Pročitaj original: arXiv AI Papers
