World Self-and-Mutual Models: Adaptive Switching Between Self-Diagnosis and Mutual Diagnosis for World-State Understanding and Model Reliability

Abstract

World models provide computational representations of environmental states, transitions, and potential outcomes. However, a world model may contain internal inconsistencies, inaccurate state estimates, incorrect transition predictions, or assumptions that no longer correspond to the environment. Diagnosing these problems is therefore essential for maintaining reliable world representations. Existing research has investigated model-based diagnosis, multi-agent fault diagnosis, self-correction, and collaborative reasoning, but these research directions do not, by themselves, establish a unified framework in which self-diagnosis and mutual diagnosis are adaptively selected as complementary operating modes within a world-model architecture. This paper proposes the World Self-and-Mutual Models (WSMM) framework, a conceptual architecture that combines world-model-based state representation with a higher-level self-and-mutual modeling principle. WSMM distinguishes two diagnostic modes. In self-diagnosis, an individual world model examines its own state estimates, transition predictions, internal constraints, and accumulated discrepancies. In mutual diagnosis, multiple world models examine one another's predictions, compare their representations, and investigate inconsistencies that cannot be adequately resolved through isolated analysis. A diagnostic controller dynamically selects between these modes according to uncertainty, prediction error, disagreement, diagnostic confidence, task risk, computational cost, and the expected value of additional evidence. The framework formalizes world-state estimation, diagnostic discrepancy, mode selection, cross-model evidence integration, and model updating. It also introduces a distinction between diagnosing the state of the external world and diagnosing the internal reliability of the model representing that world. This distinction is important because an apparent environmental anomaly may originate from a faulty model, incomplete observations, an incorrect transition assumption, or a genuine change in the environment. The proposed framework is independent of any particular model architecture and does not require large language models. It can be implemented using predictive state-space models, robotics models, simulation systems, scientific models, or heterogeneous computational representations. Finally, the paper develops research questions, testable hypotheses, evaluation metrics, baseline comparisons, and a reproducible experimental design. The proposed framework is conceptual and has not yet been empirically validated; its principal contribution is a formal research direction for adaptive switching between individual and cross-model diagnosis in world-model systems.