Improving the Reliability of Large Language Models: A Hypothesis-Driven Statistical Evaluation of Hallucination Detection and Mitigation Techniques

Large Language Models (LLMs) such as GPT-4, LLaMA, and Claude achieve remarkable fluency, yet their tendency to hallucinate — producing plausible but factually incorrect content — remains a central obstacle to reliable deployment in high-stakes domains. This paper organizes a comparative evaluation of six hallucination detection techniques (SelfCheckGPT, FActScore, Chain-of-Verification, internal-state probing, semantic entropy, and Retrieval-Augmented Generation) and six mitigation techniques around two falsifiable statistical hypotheses: that mean detection accuracy (AUROC) differs across technique families (H1), and that mean mitigation effectiveness differs across technique families (H2). A pilot dataset (n=12 per technique) built from benchmark-aggregated literature values was analyzed using one-way ANOVA, Levene's test, Shapiro–Wilk normality testing, Tukey's HSD post-hoc test, and eta-squared effect size. Both omnibus ANOVAs were significant (detection: F(5,66)=43.99, p<0.001, η²=0.77; mitigation: F(5,66)=55.24, p<0.001, η²=0.81), rejecting both nulls. Tukey HSD resolved detection into three statistically distinct tiers led by RAG Grounding Consistency and Semantic Entropy, and resolved mitigation into a tiered structure led by RAG alone. These results motivate a modular ensemble pipeline combining lightweight detection, retrieval grounding, and selective verification as the most defensible path toward reliable LLM systems.