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

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Rawte, Vipula, Roy, Rajarshi, Singh, Gurpreet, Khanna, Danush, Narsupalli, Yaswanth, Ghosh, Basab, Gupta, Abhay, Samanta, Argha Kamal, Shingote, Aditya, Vikram, Aadi Krishna, Jain, Vinija, Chadha, Aman, Sheth, Amit, Das, Amitava
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909773524369408
author Rawte, Vipula
Roy, Rajarshi
Singh, Gurpreet
Khanna, Danush
Narsupalli, Yaswanth
Ghosh, Basab
Gupta, Abhay
Samanta, Argha Kamal
Shingote, Aditya
Vikram, Aadi Krishna
Jain, Vinija
Chadha, Aman
Sheth, Amit
Das, Amitava
author_facet Rawte, Vipula
Roy, Rajarshi
Singh, Gurpreet
Khanna, Danush
Narsupalli, Yaswanth
Ghosh, Basab
Gupta, Abhay
Samanta, Argha Kamal
Shingote, Aditya
Vikram, Aadi Krishna
Jain, Vinija
Chadha, Aman
Sheth, Amit
Das, Amitava
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2507_02949
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle RADIANT: Retrieval AugmenteD entIty-context AligNmenT -- Introducing RAG-ability and Entity-Context Divergence
Rawte, Vipula
Roy, Rajarshi
Singh, Gurpreet
Khanna, Danush
Narsupalli, Yaswanth
Ghosh, Basab
Gupta, Abhay
Samanta, Argha Kamal
Shingote, Aditya
Vikram, Aadi Krishna
Jain, Vinija
Chadha, Aman
Sheth, Amit
Das, Amitava
Computation and Language
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.
title RADIANT: Retrieval AugmenteD entIty-context AligNmenT -- Introducing RAG-ability and Entity-Context Divergence
topic Computation and Language
url https://arxiv.org/abs/2507.02949