Reducing Hallucinations in Large Language Models: Abstention, Grounding, Failure-Driven Correction, and Evidence from a Small-Model Pilot Study

Large language models can produce fluent and persuasive responses even when the underlying information is incorrect, unsupported, outdated, or impossible to determine from the available evidence. This paper examines hallucination as a system-level reliability problem and reviews complementary mitigation approaches including abstention-aware post-training, uncertainty estimation, retrieval-augmented generation, grounding, verification, and failure-driven correction. The paper also reports an exploratory pilot study using a locally executed Gemma3-1B-IT model. Ten adversarial prompts were designed to test false-premise acceptance, unsupported inference, fabricated citations, unknowable facts, and future-value prediction. The pilot highlights several failure modes, including cases in which the model initially recognized uncertainty but subsequently generated unsupported information. The work argues that hallucination reduction should be evaluated together with factual accuracy, correct abstention, false-premise resistance, grounding fidelity, answer coverage, and generation stability. This manuscript is an independent research preprint and has not undergone peer review.