Multi-Criteria Evaluation of Hierarchical Reasoning, Self-Correction, and Factual Consistency in Large Language Models across Complex Language Tasks

Abstract

The rapid proliferation of large language models has necessitated the development of robust evaluation frameworks that extend beyond simple accuracy metrics. This paper introduces a comprehensive multi-criteria evaluation methodology designed to assess the capabilities of these advanced computational architectures in handling complex language tasks. Specifically, the study focuses on three foundational dimensions: hierarchical reasoning, self-correction mechanisms, and factual consistency. By systematically isolating these dimensions, the research provides a nuanced understanding of how models parse intricate problem structures, dynamically revise their internal states upon detecting errors, and maintain fidelity to established external knowledge bases. The proposed framework employs novel mathematical formulations to quantify these qualitative traits, enabling a rigorous, quantitative benchmarking process. Through extensive empirical analysis across diverse datasets, the findings reveal critical trade-offs between a model's ability to engage in deep hierarchical reasoning and its capacity to remain factually grounded. Furthermore, the evaluation of self-correction capabilities highlights persistent vulnerabilities in unsupervised revision protocols. This study contributes to the broader discourse on artificial intelligence reliability and safety by offering a structured approach to diagnosing model deficiencies, ultimately guiding the design of more resilient and dependable language processing systems