Adaptive Self-Reinforcement and Mutual Reinforcement in World Self-and-Mutual Models: A Unified Architecture for Switching Between Independent and Interdependent World-Model Evolution

Abstract

World models provide internal representations of environmental states, dynamics, and potential future outcomes. However, improving a world model can follow different learning regimes. A model may improve independently through its own prediction errors, internal consistency checks, counterfactual simulation, and interaction history. Alternatively, multiple world models may improve through mutual evaluation, information exchange, cross-model prediction, and reciprocal correction. Permanently relying on either regime can be suboptimal: independent learning may preserve blind spots, whereas continuous interaction may introduce communication costs, correlated errors, negative transfer, and unnecessary coordination. This paper proposes the World Self-and-Mutual Model framework (WSMM), a conceptual architecture for adaptive switching between self-reinforcement and mutual reinforcement at the world-model level. WSMM is not defined as a large language model, an ensemble of language models, or a conventional multi-agent prompting system. Instead, it is a model-level architecture in which one or more world models maintain internal state-transition representations, evaluate their own predictive performance, and selectively interact with other world models when reciprocal information is expected to improve model quality. The framework distinguishes self-reinforcement, in which a world model updates using internally generated or locally available learning signals, from mutual reinforcement, in which two or more world models exchange predictions, uncertainty estimates, counterfactual outcomes, or correction signals. A switching controller estimates the relative expected utility of the two modes using prediction error, uncertainty, model disagreement, information gain, transfer risk, computational cost, and communication overhead. A regularized objective combines predictive accuracy, decision usefulness, long-horizon consistency, robustness, and resource efficiency. The paper formalizes the architecture, introduces a switching algorithm, identifies stability and safety constraints, and proposes an empirical evaluation protocol. The central hypothesis is that adaptive switching can achieve a more favorable accuracy–cost–robustness trade-off than permanently independent or permanently collaborative world-model learning. This claim remains a testable hypothesis rather than an established empirical result. The framework provides a foundation for studying world-model evolution through the dynamic relationship between self-directed improvement and reciprocal improvement.