Adaptive Self–Mutual Correction Between Self-and-Interactive Models: A General Framework for Error Detection, Correction, and Mode Switching

Abstract

Correction is a fundamental requirement for intelligent systems that must maintain consistency, adapt to changing environments, and improve their internal representations. Existing approaches investigate error correction through self-correction, feedback control, iterative refinement, peer interaction, distributed systems, and collaborative reasoning. However, a general framework for self-and-interactive models that adaptively switches between internal self-correction and cross-model mutual correction remains a research direction that warrants systematic formalization and empirical investigation. This paper proposes the Adaptive Self–Mutual Correction Framework (ASMCF), a model-level framework in which multiple self-and-interactive models can correct their own internal states, representations, predictions, decisions, or behaviors, or correct one another through structured interaction. The framework does not assume that the underlying models are large language models, neural networks, or agents of any particular implementation. Instead, it treats models as general computational entities capable of representing states, generating outputs, evaluating discrepancies, and applying corrective operations. The central principle is adaptive correction-mode selection. Rather than permanently adopting self-correction or mutual correction, the framework selects a correction mode according to estimated error, uncertainty, correction effectiveness, interaction cost, model independence, and environmental conditions. It also supports alternating correction, simultaneous correction, parallel candidate generation, disagreement-triggered interaction, and return to self-correction after external feedback has been incorporated. We formalize the framework, define its principal components, introduce a correction-mode selection mechanism, describe an operational algorithm, and establish research hypotheses and an evaluation methodology. We further discuss convergence, error propagation, correlated errors, computational costs, and conditions under which mutual correction may be inferior to self-correction. The contribution is a general conceptual and methodological framework for investigating adaptive transitions between self-correction and mutual correction across heterogeneous computational systems. Its novelty relative to existing research must be established through systematic literature comparison and empirical validation rather than assumed in advance.