<![CDATA[The largest failure modes of LLMs to-date, bias and hallucination, have measurable harms in contexts where factuality and fairness are paramount. Both areas have experienced significant research growth; however, prior work on each generally operates as a disparate body of research, and there is a gap in a methodological framework for jointly measuring, tracing, and reducing both under the same experimental conditions. We provide that framework through an empirical evaluation (not a survey) of bias propagation and hallucination generation on four illustrative domains (medical, legal, finance, human resources) through a framework that addresses the three research questions: how can bias and hallucination be measured simultaneously through a replicable, domain-specific protocol; which techniques yield statistically meaningful improvements and a consistency of effectiveness; and how do causally informed methods fare against retrieval methods when tested for factual error reduction. We report new experiments using the GPT-4, LLaMA-2, and Falcon-7B models on the MIMIC-III, CrowS-Pairs, Yahoo Finance Q3 and XNLI-HR benchmarks while keeping our prompts uniform and our random seeds fixed. Methods included structural causal modeling, retrieval-augmented generation, uncertainty-aware RLHF, and hallucination-specific fine-tuning, with experiments on each method separately before merging them into combined frameworks. We observe that RAG achieved a 45% reduction in hallucination rates and that our causally guided active learning method reduced bias disparity by 25%; together, they substantially outperform either method alone. This contributes to a repeatable method for auditing bias and hallucinations, helping ensure alignment with EU AI Act standards and similar requirements.]]>
Copyrights © 2026
