Adaptive Control of Cognitive Load in LLM Interaction and the Evolution of Voice-Based LLM Systems: Toward Cognitive Interfaces for Efficient Human–AI Intellectual Collaboration

AbstractThe increasing capabilities of large language models (LLMs) have accelerated information retrieval, reasoning, text generation, and task execution, substantially expanding the productivity of human intellectual activity. At the same time, humans face new cognitive demands when understanding, evaluating, and integrating large quantities of LLM-generated information and controlling the overall activity. In particular, chat interfaces based on successive cycles of document generation and reading require substantial information-processing and task-management abilities on the part of the user.This paper treats cognitive load during LLM use as a variable to be controlled within human–AI collaborative systems. It proposes a cognitively adaptive interface designed to improve the quality of intellectual outcomes, collaborative efficiency, understanding, human agency, and sustainability.The proposal takes voice-based LLM interaction as an important technical foundation. It estimates users' comprehension and cognitive load from multimodal information, including speech, facial expressions, behavior, and dialogue history, and dynamically adjusts explanations, the amount of information presented, conversational pace, and the division of labor between humans and AI. It also introduces a parallel meta-dialogue mechanism through which users can take an overview of their primary intellectual activities and exchange assessments and impressions.The paper distinguishes adaptive cognitive-load control from the simplification of intellectual activity and presents design principles that enable diverse users to employ powerful LLM capabilities efficiently and sustainably. It also identifies technical continuity with educational LLM development and discusses applications spanning early childhood and primary education through advanced research. AI Assistance DisclosureGenerative AI was used as an auxiliary tool in drafting, adjusting expressions, and organizing the structure of this paper. The author bears final responsibility for the text, claims, citations, and factual verification. KeywordsLarge language models, Cognitive load, Human–AI collaboration, Voice interaction, Multimodal perception, Adaptive interfaces, Metacognitive support, Intellectual productivity