Análise de Alternativas para Mitigar Alucinações em Modelos de Linguagem de Larga Escala

Abstract

Large Language Models (LLMs) have become central to applications requiring natural language generation, problem solving, and decision support. Despite recent advances, these models remain susceptible to the phenomenon of hallucinations, characterized by the production of incorrect or unverifiable responses, which limits their use in sensitive domains. This work investigates different hallucination-mitigation strategies applied to the Llama3.2 3B model, evaluating approaches based on information retrieval (RAG), reranking (MMR and neural reranking), internal verification (Chain-of-Verification), and post-editing with a reviewer agent (Answer + Reviewer). Standardized experiments were conducted on the TruthfulQA and ARC Challenge benchmarks, which evaluate accuracy and reasoning ability, respectively. The results indicate that information-retrieval techniques yield relevant gains in the ARC Challenge—particularly RAG with reranking. Conversely, the internal verification technique, represented by CoVe, achieved remarkably strong results on the TruthfulQA compared with the baseline model. Therefore, we conclude that effective hallucination mitigation depends on the nature of the task and that hybrid combinations of retrieval and internal verification represent a promising direction for developing more reliable language models.