Work

Adaptive Mutual-Human Selection: A Unified Framework for Multi-LLM Collaborative Selection, Independent Human Selection, and Context-Driven Adaptive Switching

Year: 2026
Type: preprint
Abstract: Selection is a fundamental function underlying artificial intelligence, autonomous agents, scientific research, education, organizational management, decision support, and human problem-solving. Existing computational selection approaches commonly rely on individual model predictions, predefined ranking algorithms, ensemble aggregation, or multi-agent collaboration. Human selection, meanwhile, relies on individual or collective human judgment, experience, preferences, and contextual understanding. Although these approaches provide complementary capabilities, a general architecture that enables a system to adaptively determine whether selection should be performed through collaboration among multiple large language models (LLMs), independent human selection, or a combination of both remains an important theoretical direction. This paper introduces Adaptive Mutual-Human Selection (AMHS), a function-specific theoretical framework developed from the Universal Mutual-Human Model (UMHM). The proposed framework establishes two principal selection mechanisms: Mutual Selection and Human Selection. Mutual Selection enables multiple independent or specialized LLMs to collaboratively examine alternatives, exchange information, compare candidate options, challenge recommendations, evaluate evidence, and reach a collective selection outcome. Human Selection enables an individual or group of humans to independently perform the selection function without requiring multiple AI participants. Unlike conventional architectures that permanently assign selection authority to either automated systems or humans, AMHS introduces Adaptive Selection Mechanism Selection, Adaptive Selection Switching, Adaptive Selection Combination, and Adaptive Selection Reconfiguration. The architecture determines which selection mechanism is appropriate according to task characteristics, model reliability, uncertainty, disagreement, human expertise, contextual requirements, resource constraints, and decision consequences. A formal mathematical representation, adaptive selection policy, transition mechanism, decision-theoretic objective, architectural design, theoretical propositions, application scenarios, governance principles, and empirical evaluation methodology are developed. The framework explicitly distinguishes collaborative selection from human selection and treats the active mechanism as a dynamic operational decision rather than a permanent system property. To the best of our knowledge, this paper is the first to formally articulate this particular UMHM-grounded, function-general integration of multi-LLM collaborative selection and independent human selection through an explicit architecture for adaptive mechanism selection, switching, combination, and reconfiguration. This originality claim concerns the specific theoretical integration and does not claim to originate ensemble learning, multi-agent decision-making, human decision-making, or mixed-initiative systems. The proposed framework provides a conceptual foundation for intelligent systems capable of changing how selection is performed according to the requirements of each task and the complementary capabilities of artificial and human participants.
Language: English
Open Access status: green
OpenAlex ID: W7236864094
Cites:
Cited by:
Related to: 10