Adaptive Self-Play and Multi-Model Play in Self-and-Interaction Models: A Unified Framework for Dynamic Switching Between Internal and Inter-Model Strategic Exploration

Abstract

Self-play and multi-model interaction provide different mechanisms for exploring complex problem spaces. Self-play enables a model to generate, evaluate, and challenge its own strategic alternatives, whereas multi-model play enables multiple models to interact through competition, cooperation, adversarial examination, negotiation, or other forms of strategic engagement. However, permanently adopting either mode may impose limitations. Self-play may suffer from insufficient diversity, persistent internal biases, or limited access to alternative perspectives. Multi-model play may introduce communication costs, coordination overhead, correlated errors, and unproductive interactions. This paper proposes an adaptive self-play and multi-model play framework for Self-and-Interaction Models (SIMs), a model-level conceptual architecture distinct from any particular large language model or conventional multi-agent implementation. The proposed framework dynamically selects between internal strategic exploration and inter-model strategic interaction according to uncertainty, disagreement, expected information gain, exploration progress, computational cost, and task requirements. A unified decision mechanism estimates the relative utility of both modes and supports transitions between them, including repeated switching when the problem state changes. The framework further introduces model-level game representations, switching hysteresis, interaction budgets, strategic diversity preservation, and mode-specific evaluation. A mathematical formulation, an operational algorithm, a complexity discussion, and a falsifiable experimental design are presented. This paper establishes a conceptual and methodological foundation rather than claiming experimentally demonstrated superiority. Its potential contribution lies in treating self-play and multi-model play as complementary, adaptively selectable operating modes within a unified model-level architecture.