Adaptive Switching Between Parallel-World Self-Correction and Mutual Correction for Multiple Large Language Models

Abstract

Large language models (LLMs) can generate incorrect, incomplete, or internally inconsistent responses. Self-correction allows a model to inspect and revise its own outputs, whereas mutual correction enables multiple model instances or heterogeneous models to identify and address one another's errors. However, relying exclusively on either mechanism may limit the diversity of corrective perspectives, increase computational costs, or reinforce existing mistakes. This paper proposes Adaptive Parallel-World Self-Correction and Mutual Correction (APSMC), a unified framework in which multiple LLMs dynamically switch between independent self-correction and collaborative mutual correction according to the observed characteristics of their outputs and the estimated benefits of further interaction. The framework introduces parallel worlds, defined as independently maintained corrective trajectories generated by the same underlying model under different correction objectives, prompts, reasoning constraints, or contextual perspectives. Each trajectory performs self-correction before selectively exchanging information with other trajectories. An adaptive controller determines whether the system should continue independent correction, initiate pairwise mutual correction, activate group-level correction, or terminate the process. The framework also supports heterogeneous LLMs that participate in correction without belonging to the same parallel-world group. Unlike fixed self-correction and permanently collaborative architectures, APSMC treats the choice between independent and collaborative correction as a dynamic decision-making problem. This paper formalizes the framework, presents a correction-selection algorithm, defines evaluation metrics, and proposes an experimental protocol for assessing accuracy, correction reliability, diversity, computational efficiency, and resistance to collective errors. The framework is a research proposal; its effectiveness and novelty relative to existing methods require systematic empirical validation.