Work

Adaptive Self–Human Verification: A Unified Framework for Dynamic Verification Selection, Switching, and Collaborative Validation in Large Language Models

Year: 2026
Type: preprint
Abstract: Large Language Models (LLMs) increasingly perform tasks that require the verification of generated information, reasoning processes, factual claims, computational results, decisions, and other model-produced outputs. Verification is essential for improving reliability, identifying errors, reducing hallucinations, and establishing confidence in AI-generated results. Existing research has investigated model-based self-verification, self-consistency, iterative self-evaluation, external verification, human fact-checking, and human-in-the-loop assessment. However, these approaches frequently emphasize particular verification techniques or fixed arrangements of autonomous and human participation. A broader architectural question remains insufficiently formalized: should an LLM always verify its own outputs, should verification always be delegated to humans, or should the verification responsibility dynamically move between these alternatives according to the characteristics and requirements of each task? This paper introduces the Adaptive Self–Human Verification Framework (ASHVF), a proposed mechanism-independent framework for dynamically organizing verification in LLM-based systems. Grounded in the Universal Self–Human Model (USHM), the framework establishes two fundamental verification dimensions: Self-Verification (S), in which an LLM performs verification through its own internal or self-directed mechanisms, and Human Verification (H), in which human individuals or groups examine, assess, validate, or challenge model-generated results. These dimensions produce three non-empty operational configurations: Self-Verification, Human Verification, and Self–Human Verification. The central contribution is an adaptive verification architecture that enables systems to select, switch between, combine, and reconfigure verification mechanisms rather than permanently relying on a single verification pathway. The proposed framework introduces Adaptive Verification Selection, Adaptive Verification Switching, Adaptive Verification Combination, and Adaptive Verification Reconfiguration. It further formalizes verification capability, uncertainty, verification risk, human expertise, decision authority, verification cost, and switching conditions within a unified decision model. Importantly, the framework does not assume that model self-verification is inherently reliable or that human verification is invariably correct. Instead, it treats both as potentially fallible verification sources whose suitability must be assessed according to task-specific evidence and operational constraints. To the best of our knowledge, this paper is the first to formally propose a unified, mechanism-independent LLM verification architecture specifically organized around adaptive transitions between model self-verification and human verification as distinct but dynamically combinable verification dimensions. This priority claim concerns the particular theoretical integration and adaptive architecture developed in this paper, rather than the broader established fields of self-verification, automated fact-checking, human-in-the-loop verification, or human–AI collaboration. The paper establishes a conceptual foundation for developing intelligent systems that determine not only what should be verified, but also who should verify it, when verification responsibility should change, and how autonomous and human verification should be coordinated.
Language: English
Open Access status: green
OpenAlex ID: W7236569569
Cites:
Cited by:
Related to: 10