Exécution de notre outil de synthèse sur un nouvel article

Résumé : Large language models exhibit remarkable performance on knowledge-intensive tasks yet remain vulnerable to hallucination, producing plausible but factually unsupported content. Retrieval-Augmented Generation (RAG) partially addresses this limitation by grounding model outputs in retrieved documents, but conventional RAG pipelines neither assess retrieval quality nor verify output reliability. This paper introduces a dual-component enhancement to the standard RAG paradigm. The first component, Confidence-Aware Retrieval, combines dense semantic similarity with BM25 lexical scores into a hybrid confidence measure, filtering retrieved passages below a tunable threshold. The second component, Self-Consistency Verification, generates multiple candidate responses and accepts a final answer only when inter-response semantic agreement exceeds a defined threshold, otherwise triggering query expansion and re-retrieval. Evaluated on 500 samples from the HaluEval QA benchmark, the integrated system achieves a hallucination rate of 23.6% compared to 42.5% for vanilla RAG—representing a 44.5% relative reduction—while improving faithfulness from 0.520 to 0.730. These results demonstrate that jointly optimising retrieval confidence and generation consistency constitutes a practical and effective hallucination mitigation strategy.