Adaptive Mutual-Human Collaborative Learning: A Universal Framework for Dynamic Switching Between Multi-Model Learning and Human Learning
Year: 2026
Type: preprint
Abstract: The rapid development of large language models (LLMs), multi-agent artificial intelligence, and intelligent educational systems has created new possibilities for learning through interactions among multiple artificial intelligence models and human participants. Existing research has investigated collaborative learning, multi-agent systems, human learning, AI-assisted education, human–AI collaboration, self-regulated learning, and adaptive learning technologies. However, these research directions frequently emphasize particular learning arrangements, fixed participation structures, or predefined forms of human–AI assistance. An important theoretical challenge remains: how can a learning system dynamically determine whether learning should be conducted through collaboration among multiple LLMs, through human learning, or through a combination of both, and how can it adaptively switch between these mechanisms as learning requirements change? This paper introduces Adaptive Mutual-Human Collaborative Learning (AMHCL), a learning-specific theoretical framework developed from the Universal Mutual-Human Model (UMHM). The framework establishes three fundamental learning configurations: Mutual Learning, Human Learning, and Mutual–Human Collaborative Learning. Mutual Learning represents learning-related interaction among multiple LLMs or other artificial intelligence models, including knowledge exchange, reciprocal explanation, collaborative knowledge construction, cross-model correction, knowledge integration, and distributed learning. Human Learning represents the independent or collective acquisition, interpretation, organization, retention, reflection, and application of knowledge by human individuals or groups. Mutual–Human Collaborative Learning combines these two dimensions, enabling humans and multiple AI models to participate in coordinated learning processes. The central contribution is an adaptive mechanism that selects, switches, combines, and reconfigures these learning configurations according to knowledge gaps, learning objectives, performance, uncertainty, model capabilities, human learning needs, cognitive conditions, resource availability, and contextual requirements. Unlike architectures that permanently assign learning to either AI systems or human participants, AMHCL treats the learning mechanism itself as an adaptive decision. The paper formalizes the dimensional architecture, learning-state representation, adaptive selection policy, switching conditions, collaborative knowledge integration, learning feedback mechanisms, and evaluation methodology. It further distinguishes improvements in AI model knowledge or capabilities from genuine human learning outcomes, preventing the mistaken assumption that improved AI performance automatically means improved human understanding. To the best of our knowledge, this paper is the first to formally articulate, within the UMHM theoretical lineage, a function-specific architecture that integrates multi-model collaborative learning and human learning through explicit adaptive selection, switching, combination, and structural reconfiguration. This contribution is narrowly defined and does not claim that collaborative learning, multi-agent learning, or human–AI co-learning originated with this work. The framework provides a theoretical foundation for adaptive learning systems in which artificial and human learning processes can operate independently, collaboratively, sequentially, or dynamically according to the requirements of each learning situation.
Author: Zhang, Jincheng
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7236582976
