Exécution de notre outil de synthèse sur un nouvel article

Résumé : Large Language Models (LLMs) have demonstrated remarkable fluency but remain fundamentally unreliable for knowledge-intensive tasks due to hallucination, temporal staleness, and lack of verifiability.Retrieval-Augmented Generation (RAG) has emerged as a dominant architectural paradigm that grounds generation in external evidence.This paper provides a comprehensive investigation into how RAG reshapes LLM reliability across factual accuracy, citation fidelity, temporal freshness, and robustness to noise.We benchmark five RAG variants against closed-book baselines using secondary datasets including Natural Questions, TriviaQA, HotpotQA, and RGB.Results show an average 23.7% reduction in hallucination rate and 31.2%improvement in citation grounding when using advanced retrieval with self-reflection.However, we identify critical failure modes: retriever-generator misalignment, context distraction, and over-reliance on irrelevant passages.We propose a reliability framework and demonstrate that corrective and modular RAG designs significantly mitigate these risks.Our findings establish RAG as necessary but not sufficient for trustworthy deployment.