Large Language Models (LLMs) demonstrate strong generative capability but remain prone to hallucination, the production of fluent yet factually incorrect or unverifiable content. Small Language Models (SLMs), typically spanning a few hundred million to roughly eight billion parameters, have emerged as a practical route to more controllable and trustworthy generation. This paper reviews the architecture and parameter scale of representative SLM families, formally defines hallucination and its sub-types, and proposes reference workflow architectures — a training/deployment pipeline, a retrieval-grounded inference workflow, and a layered mitigation stack — that explain how SLMs reduce hallucination in practice. Techniques including curated-data training, knowledge distillation, retrieval-augmented generation (RAG), instruction tuning, reinforcement learning from human feedback (RLHF) / direct preference optimisation (DPO), and constrained decoding are analysed with respect to the specific hallucination pattern each addresses. Results synthesised from the literature indicate that hallucination reduction in SLMs is best understood as a systems-level property arising from disciplined data curation and layered verification, rather than a consequence of parameter count alone.
Prateek Dutta (Thu,) studied this question.