Beyond Autonomous Planning: An Adaptive Self–Human Planning Framework for Dynamic Selection, Switching, Collaboration, and Reconfiguration in Large Language Models
Large language models (LLMs) and intelligent agents are increasingly capable of generating plans, decomposing complex objectives, organizing subtasks, allocating resources, predicting future states, and coordinating multi-step operations. Nevertheless, autonomous planning is not universally appropriate. Certain tasks can be planned efficiently through a model's own internal mechanisms, whereas others require human knowledge, contextual understanding, strategic judgment, practical experience, or direct human authority. A further category involves planning processes in which artificial intelligence and humans contribute complementary capabilities. Existing research has investigated autonomous planning, LLM-based task planning, human-in-the-loop planning, mixed-initiative systems, interactive planning, and human–AI collaboration. However, these research directions do not, by themselves, establish a general mechanism-independent architecture in which self-directed planning and human-performed planning constitute distinct but dynamically interchangeable functional configurations, with adaptive switching between them as a central planning principle. This paper introduces the Adaptive Self–Human Planning Framework (ASH-PF), a function-specific theoretical architecture grounded in the Universal Self–Human Model (USHM). The framework establishes three fundamental planning configurations: Self-Planning (SP), in which an intelligent system independently formulates and organizes plans; Human Planning (HP), in which human individuals or groups perform the planning function; and Self–Human Collaborative Planning (SHCP), in which artificial intelligence and humans jointly contribute to planning. These configurations are not permanent system classifications or mutually exclusive planning philosophies. Instead, they constitute dynamically selectable operational arrangements that may change before, during, and across planning activities. The framework introduces five integrated capabilities: Universal Planning Functional Generality, Adaptive Planning Configuration Selection, Adaptive Planning Mechanism Switching, Adaptive Planning Combination, and Adaptive Planning Reconfiguration. A formal model is developed to represent planning objectives, task complexity, uncertainty, system capability, human expertise, resource availability, authority constraints, switching costs, and planning outcomes. The framework further distinguishes planning configuration changes from ordinary plan revision, thereby separating changes in who performs planning from changes in what a plan contains. It also addresses human authority, planning accountability, uncertainty, automation bias, human workload, and safety constraints. To the best of our knowledge, this paper is the first to formally articulate the specific mechanism-independent, function-general planning architecture that unifies self-directed planning and human-performed planning through adaptive configuration selection, switching, combination, and reconfiguration as its central organizing principle. This priority claim concerns the particular theoretical integration presented here, not the broader established fields of autonomous planning, human planning, mixed-initiative planning, or human–AI collaboration. The proposed framework establishes a conceptual foundation for intelligent systems capable of dynamically determining whether planning should be performed independently, by humans, or through an adaptive combination of both.
Authors
- Jincheng Zhang (ORCID: https://orcid.org/0009-0005-1860-0009)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24359898
- Primary Topic
- AI-based Problem Solving and Planning
- Type
- preprint