Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper presents a systematic review and meta-analysis of 98 publications from 2023-2025 to analyze the architectures and performance of generative AI-powered teaching assistants, focusing on Retrieval-Augmented Generation (RAG) systems. The study examines the fusion of Transformer-based LLMs and RAG across theory, architecture, mechanism, and application, identifying key technical improvement directions like domain knowledge base construction and hybrid retrieval optimization. Meta-analysis reveals that RAG-enhanced systems achieve significantly higher accuracy (87.3%) and learning effectiveness (Cohen's d = 0.68) compared to pure generative models, with Transformer and RAG integration becoming dominant architectures.
RAG-enhanced question-answering systems beat pure generative models by 16% in accuracy, but still face challenges in answer accuracy, scenario adaptation, and academic integrity.
Teacher education confronts the contradiction between personalization and scalability. This paper employs a systematic review methodology to examine 98 qualified publications retrieved from 2023 to 2025 through PRISMA protocols from databases including CNKI, Wanfang Data, and Web of Science. The inclusion criteria encompassed studies explicitly addressing generative AI applications in educational question-answering scenarios, containing technical architectures or empirical data, and published in peer-reviewed journals or conferences. The analysis examines generative AI-powered teaching assistants across four dimensions: theory, architecture, mechanism, and application. This study provides an in-depth analysis of the fusion mechanisms between Transformer-based large language models and Retrieval-Augmented Generation (RAG) technologies in intelligent question-answering systems, encompassing sparse-dense hybrid retrieval strategies, multi-turn dialogue state management algorithms, and machine learning models for personalized recommendations. Results demonstrate that systems based on large language models and RAG exhibit advantages in knowledge transmission, capability cultivation, and learning support, yet remain constrained by answer accuracy, scenario adaptation, and academic integrity concerns. This research systematically traces the technological evolution of educational question-answering systems across four generations, from expert systems to large language models, and proposes key technical improvement directions including domain knowledge base construction, hybrid retrieval optimization, and multimodal information processing. Meta-analysis of 42 empirical studies demonstrates that RAG-enhanced systems achieve 87.3% accuracy (versus 71.5% for pure generative models) and Cohen's d = 0.68 for learning effectiveness, while architectural adoption exhibits exponential growth with Transformer and RAG integration reaching 96% and 81% respectively by early 2025. The results offer references for architectural design and performance evaluation for intelligent computing systems application in the vertical field of education, promoting further deep integration of computational intelligence and artificial applications in education.