Exploration on Cognitive Processing Mechanisms of Human-machine Collaborative Learning Supported by Multi-agents

With the rapid development of generative artificial intelligence (GenAI), multi-agents that integrate large language models with virtual digital humans have gradually evolved into the important collaborative entities for human-machine collaborative learning. However, the intrinsic mechanisms for how learners and multi-agents carry out effective collaboration remains unclear. Accordingly, this paper conducts an experiment on multi-agents human-machine collaborative learning in the “German Speech” course. Thirty German language learners are recruited from Z University to complete two types of human-machine collaborative learning tasks involving declarative knowledge and procedural knowledge, and their electroencephalogram (EEG) signals and human-machine dialogue text data are collected simultaneously. Through EEG analysis and latent dirichlet allocation (LDA) topicanalysis, this paper finds that the main effect of brain regions is significant, and brain-wave power of each frequency band shows certain regional differences. The human-machine dialogue tests are stably clustered into two or three topics: declarative knowledge learning focuses on knowledge interpretation and value judgment, while procedural knowledge learning involves more comparative analysis, meaning construction, and integrated application. Based on this conclusion, this paper constructs an alignment framework for human-machine collaborative cognition and EEG supported by multi-agents, and reveals the cognitive processing mechanism of human-machine collaborative learning: multi-agents-dominated external perception at the information input phase, human-machine primary-auxiliary collaborative meaning construction at the content-comprehension phase, human-machine-jointly dominated information integration at the structural-organization phase, and learner-dominated metacognitive activation at the regulation-reflection phase. Integrating implicit EEG signals with explicit dialogue texts, this paper reveals the cognitive processing characteristics and human-machine role division of human-machine collaborative learning supported by multi-agents at different stages, and can provide references for the development of multi-agent system and the design of human-machine collaborative teaching.