Exécution de notre outil de synthèse sur un nouvel article

Résumé : Retrieval-Augmented Generation (RAG) systems have significantly improved the factual grounding of large language models (LLMs).However, challenges remain in ensuring both semantic correctness and transparent attribution, as models may still produce hallucinated or unverified outputs.This paper proposes the Double-Lock Framework, a multi-layered architecture that integrates data engineering, linguistic attribution modeling, and dual validation mechanisms to mitigate hallucinations.A high-fidelity "Gold Standard" dataset is constructed using strict filtering criteria based on factual consistency and hallucination labels.The framework introduces attribution markers to enhance interpretability and employs both supervised finetuning and reward-based optimization for model training.A two-stage validation mechanism combining regex-based linguistic checks and predictive AI scoring ensures both transparency and factual accuracy.Experimental evaluation on a 6,000-sample dataset demonstrates that while regex-based validation alone results in a 7.72% hallucination leakage rate, the proposed Double-Lock mechanism achieves 0% leakage.The framework effectively bridges the gap between retrieval and human-understandable knowledge synthesis, enabling the development of reliable and trustworthy AI systems.