Adaptive Switching Between Self-Verification and Mutual Verification in Multiple Large Language Models

Abstract

Large language models (LLMs) can generate plausible but incorrect responses, and their ability to verify their own outputs is limited by uncertainty, shared reasoning biases, and insufficient access to independent evidence. Self-verification allows a model to inspect its own output, whereas mutual verification enables multiple models to evaluate one another's responses. However, relying exclusively on either mode may be inefficient or unreliable: self-verification can preserve the original model's errors, while mutual verification introduces additional computational costs and may propagate correlated mistakes. This paper proposes Adaptive Self–Mutual Verification (ASMV), a framework that dynamically switches between self-verification and mutual verification according to the estimated verification requirements of each generated response. The framework consists of an initial generation stage, a self-verification stage, a verification-state estimator, an adaptive mode-selection controller, a mutual-verification stage, and a final decision module. The controller considers uncertainty, disagreement, estimated error severity, verification history, evidence availability, and computational cost to determine whether a model should continue checking its own output or request verification from other models. Rather than treating self-verification and mutual verification as mutually exclusive system designs, ASMV treats them as complementary operational modes within a unified decision process. The paper formulates the framework mathematically, specifies an implementable algorithm, develops testable research hypotheses, and proposes an evaluation protocol covering reasoning accuracy, factuality, calibration, verification efficiency, and robustness. Because the framework has not yet been empirically evaluated in this paper, its expected benefits are presented as hypotheses rather than established results. The proposed research aims to investigate whether adaptive verification can provide a more effective balance between reliability, computational expenditure, and resistance to shared model errors than fixed verification strategies.