Adaptive Self–Mutual Error Correction in World Models: A Unified Framework for State-Dependent Switching Between Internal Repair and Cross-Model Correction

Abstract

World models provide internal representations of environmental dynamics and support prediction, planning, simulation, and decision-making. However, errors in world models may persist when a system relies exclusively on its own internal representations, while external correction can introduce communication costs, conflicting evidence, correlated errors, and unnecessary coordination. This paper proposes Adaptive Self–Mutual Error Correction in World Models (ASME-WM), a conceptual framework that integrates self-models, mutual models, and world models into a unified error-correction architecture. The framework distinguishes self-correction, in which a model detects and repairs inconsistencies within its own representations, from mutual correction, in which multiple models exchange diagnostic information and evaluate one another's predictions or proposed repairs. Rather than fixing the system to either correction mode, ASME-WM selects between them according to estimated error severity, internal uncertainty, cross-model disagreement, historical correction reliability, computational cost, and the risk of propagating incorrect information. A world-model consistency layer evaluates whether candidate corrections improve predictive and structural consistency without unnecessarily destabilizing previously reliable knowledge. The proposed architecture is independent of large language models and can be implemented using learned state-space models, probabilistic models, dynamical systems, graph-based representations, model-based reinforcement learning, or hybrid computational mechanisms. This paper formalizes the architecture, defines its operational components, presents an adaptive switching algorithm, and develops theoretical propositions and an empirical evaluation protocol. The framework is presented as a research proposal rather than an experimentally validated system. Its principal research hypothesis is that state-dependent coordination between internal repair and cross-model correction can provide a more favorable balance among correction quality, robustness, computational efficiency, and model stability than fixed correction strategies.