Adaptive Self–Human Diagnosis in Large Language Models: A Universal Framework for Functional Diagnosis, Adaptive Diagnostic Selection, Dynamic Switching, Collaborative Diagnosis, and Reconfiguration
Large language models (LLMs) increasingly perform complex computational and cognitive functions, yet their ability to identify, investigate, interpret, and diagnose problems remains constrained by the reliability of their internal assessment mechanisms. When an error, abnormal condition, performance degradation, inconsistency, or unexpected behavior occurs, an intelligent system may attempt to diagnose the problem through its own internal mechanisms or request human participation. Existing research has investigated self-evaluation, self-reflection, error detection, automated fault diagnosis, human-in-the-loop systems, expert diagnosis, and human–AI collaboration. However, these approaches commonly investigate particular diagnostic tasks, technical implementations, or fixed patterns of human involvement. A general mechanism-independent architecture that treats self-directed diagnosis and human-performed diagnosis as dynamically selectable and combinable functional configurations across arbitrary diagnostic domains remains insufficiently formalized. This paper introduces the Adaptive Self–Human Diagnostic Framework (ASHDF), a universal conceptual framework for dynamically allocating diagnostic functions between self-directed mechanisms and human participation. Building upon the two-dimensional conceptual foundation of the Universal Self–Human Model (USHM), ASHDF establishes three fundamental diagnostic configurations: Self-Diagnosis (S), Human Diagnosis (H), and Self–Human Diagnosis (S–H). Self-Diagnosis enables an intelligent system to identify symptoms, investigate potential causes, analyze evidence, estimate diagnostic confidence, and formulate diagnostic conclusions through its own mechanisms. Human Diagnosis enables human individuals or groups to perform these activities through human expertise, contextual understanding, independent assessment, and professional judgment. Self–Human Diagnosis combines both sources of diagnostic capability through parallel, sequential, iterative, complementary, or conditional participation. The central contribution is the introduction of adaptive diagnostic configuration selection and switching, through which an intelligent system determines whether it should diagnose a problem independently, delegate the diagnostic function to a human, or combine both approaches. The selection process considers diagnostic uncertainty, task complexity, system capability, evidence quality, human expertise, diagnostic consequences, resource availability, and operational constraints. The framework further formalizes diagnostic decomposition, evidence acquisition, causal investigation, diagnostic confidence, disagreement analysis, adaptive escalation, transition costs, and post-diagnostic reassessment. Importantly, diagnosis is distinguished from simple error detection and correction: identifying an abnormality does not necessarily establish its underlying cause, and correcting an output does not necessarily demonstrate that the original problem has been correctly diagnosed. To the best of our knowledge, this paper is the first to formally articulate the specific mechanism-independent, function-general architecture that unifies self-diagnosis and human diagnosis through adaptive configuration selection, dynamic switching, combined diagnostic participation, and structural reconfiguration within a unified Self–Human framework. This priority claim concerns the particular theoretical integration and architecture proposed in this paper, rather than the broader established fields of automated diagnosis, fault detection, clinical diagnosis, self-evaluation, human-in-the-loop systems, or human–AI collaboration. The paper establishes a theoretical foundation for designing and evaluating intelligent systems that adaptively determine who or what should perform a diagnostic function under changing operational conditions.
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.24359820
- Primary Topic
- Engineering and Test Systems
- Type
- preprint