A COMPREHENSIVE REVIEW OF LARGE LANGUAGE MODEL EVALUATION, RELIABILITY, AND SAFETY

Abstract

Large language models (LLMs) have moved rapidly from research artefacts to deployed systems embedded in education, healthcare, software engineering, and autonomous agent pipelines, making rigorous evaluation of their capability, reliability, and safety a matter of direct practical consequence rather than purely academic interest. This paper presents a comprehensive review of the state of LLM evaluation, organised around three interlinked themes. First, it traces the evolution of AI evaluation paradigms, from static, single-task benchmarks of the 1990s and 2000s, through the deep-learning and foundation-model era, to the current agent-centric evaluation challenges posed by autonomous, tool-using LLM systems. Second, it synthesises the principal evaluation methodologies in current use, spanning automatic intrinsic benchmarks, structured human evaluation frameworks scored across dimensions such as relevance, coherence, coverage, and harm, LLM-as-a-judge automatic evaluation, and domain-specific evaluation frameworks illustrated through the case of medical LLM assessment. Third, it examines reliability and safety as distinct but related evaluation concerns, reviewing hallucination and factuality assessment, consistency and robustness testing, inter-rater and human-AI agreement studies, benchmark contamination, alignment techniques including reinforcement learning from human feedback and Constitutional AI, red-teaming methodology, and the emerging challenge of evaluating autonomous LLM-based agents. The review concludes by identifying persistent gaps, including benchmark saturation, the cost and subjectivity of human evaluation, cultural and linguistic bias in existing benchmark suites, and the absence of standardised protocols for long-horizon agentic evaluation, and by outlining practical mitigation strategies, including transparency, shared evaluation infrastructure, and broader stakeholder participation, that can support more trustworthy and reproducible evaluation practice going forward.