arXiv 2507.029492025-09-06cs.CL

RADIANT: 检索增强的实体-上下文对齐——引入RAG能力和实体-上下文发散度

RADIANT: Retrieval AugmenteD entIty-context AligNmenT -- Introducing RAG-ability and Entity-Context Divergence

AI Institute, University of South Carolina（南卡罗来纳大学人工智能研究所）；Meta；Amazon GenAI（亚马逊生成式人工智能）

Vipula Rawte, Rajarshi Roy, Gurpreet Singh, Danush Khanna, Yaswanth Narsupalli, Basab Ghosh, Abhay Gupta, Argha Kamal Samanta, Aditya Shingote, Aadi Krishna Vik… 展开作者

Vipula Rawte, Rajarshi Roy, Gurpreet Singh, Danush Khanna, Yaswanth Narsupalli, Basab Ghosh, Abhay Gupta, Argha Kamal Samanta, Aditya Shingote, Aadi Krishna Vikram, Vinija Jain, Aman Chadha, Amit Sheth, Amitava Das

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

针对LLM在检索增强生成中事实不一致的问题，提出实体-上下文发散度指标和Radiant框架，通过扩展DPO优化证据整合，提升RAG能力并减少幻觉。

摘要：

随着大型语言模型（LLMs）的持续进步，检索增强生成（RAG）已成为一种通过将外部知识整合到生成过程中来提升事实准确性的关键技术。然而，LLMs 往往无法忠实地将检索到的证据整合到其生成的响应中，导致事实不一致。为了量化这一差距，我们引入了实体-上下文发散度（ECD），这是一种衡量检索信息在模型输出中被准确反映程度的指标。我们系统地评估了当代 LLMs 在检索增强设置下保持事实一致性的能力，我们将这种能力定义为 RAG 能力（RAG-ability）。我们的实证分析表明，大多数 LLMs 的 RAG 能力仍然较低，突显了在实体保留和上下文保真度方面的重大挑战。本文介绍了 Radiant（检索增强的实体-上下文对齐），这是一个新颖的框架，将 RAG 与对齐相结合，旨在优化检索证据与生成内容之间的相互作用。Radiant 扩展了直接偏好优化（DPO），以教导 LLMs 如何将提供的附加信息整合到后续生成中。作为一种行为纠正机制，Radiant 在各种检索场景（如嘈杂的网络上下文、知识冲突和幻觉减少）中提升了 RAG 性能。这使得生成内容更加可靠、上下文有依据且事实连贯。

英文摘要：

As Large Language Models (LLMs) continue to advance, Retrieval-Augmented Generation (RAG) has emerged as a vital technique to enhance factual accuracy by integrating external knowledge into the generation process. However, LLMs often fail to faithfully integrate retrieved evidence into their generated responses, leading to factual inconsistencies. To quantify this gap, we introduce Entity-Context Divergence (ECD), a metric that measures the extent to which retrieved information is accurately reflected in model outputs. We systematically evaluate contemporary LLMs on their ability to preserve factual consistency in retrieval-augmented settings, a capability we define as RAG-ability. Our empirical analysis reveals that RAG-ability remains low across most LLMs, highlighting significant challenges in entity retention and context fidelity. This paper introduces Radiant (Retrieval AugmenteD entIty-context AligNmenT), a novel framework that merges RAG with alignment designed to optimize the interplay between retrieved evidence and generated content. Radiant extends Direct Preference Optimization (DPO) to teach LLMs how to integrate provided additional information into subsequent generations. As a behavior correction mechanism, Radiant boosts RAG performance across varied retrieval scenarios, such as noisy web contexts, knowledge conflicts, and hallucination reduction. This enables more reliable, contextually grounded, and factually coherent content generation.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。