Exécution de notre outil de synthèse sur un nouvel article

Résumé : Large language models (LLMs) have demonstrated exceptional capabilities across natural language processing (NLP) tasks; however, they remain persistently susceptible to generating factually incorrect content—a phenomenon broadly termed hallucination. Retrieval-augmented generation (RAG) has emerged as a principled paradigm for mitigating this limitation by grounding model outputs in dynamically retrieved external evidence, thereby substantially reducing factual errors in knowledge-intensive settings. This paper presents a comprehensive review of RAG research, tracing developments from early retrieval-enhanced pretraining frameworks to adaptive and self-reflective architectures. We examine how retrieval strategies including dense passage retrieval (DPR), sparse retrieval, and hybrid methods interact with generative components to suppress hallucination. We analyze the Knowledge-Intensive Language Tasks (KILT) benchmark and open-domain question answering (QA) datasets as primary evaluation vehicles, synthesizing empirical evidence demonstrating that RAG consistently lowers factual error rates relative to purely parametric LLMs. We further discuss challenges including retrieval quality, knowledge conflict resolution, multi-hop reasoning, and domain adaptation, and outline future directions essential for realizing the full potential of RAG in high-stakes natural language generation (NLG) applications.