Enregistré dans:
Détails bibliographiques
Auteurs principaux: Ghosh, Reshmi, Seetharaman, Rahul, Wadhwa, Hitesh, Aggarwal, Somyaa, Basu, Samyadeep, Srinivasan, Soundararajan, Zhao, Wenlong, Chaudhari, Shreyas, Aghazadeh, Ehsan
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:https://arxiv.org/abs/2410.00857
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917792726384640
author Ghosh, Reshmi
Seetharaman, Rahul
Wadhwa, Hitesh
Aggarwal, Somyaa
Basu, Samyadeep
Srinivasan, Soundararajan
Zhao, Wenlong
Chaudhari, Shreyas
Aghazadeh, Ehsan
author_facet Ghosh, Reshmi
Seetharaman, Rahul
Wadhwa, Hitesh
Aggarwal, Somyaa
Basu, Samyadeep
Srinivasan, Soundararajan
Zhao, Wenlong
Chaudhari, Shreyas
Aghazadeh, Ehsan
contents Retrieval Augmented Generation (RAG) is a widely used approach for leveraging external context in several natural language applications such as question answering and information retrieval. Yet, the exact nature in which a Language Model (LM) leverages this non-parametric memory or retrieved context isn't clearly understood. This paper mechanistically examines the RAG pipeline to highlight that LMs demonstrate a "shortcut'' effect and have a strong bias towards utilizing the retrieved context to answer questions, while relying minimally on model priors. We propose (a) Causal Mediation Analysis; for proving that parametric memory is minimally utilized when answering a question and (b) Attention Contributions and Knockouts for showing the last token residual stream do not get enriched from the subject token in the question, but gets enriched from tokens of RAG-context. We find this pronounced "shortcut'' behaviour to be true across both LLMs (e.g.,LlaMa) and SLMs (e.g., Phi)
format Preprint
id arxiv_https___arxiv_org_abs_2410_00857
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Quantifying reliance on external information over parametric knowledge during Retrieval Augmented Generation (RAG) using mechanistic analysis
Ghosh, Reshmi
Seetharaman, Rahul
Wadhwa, Hitesh
Aggarwal, Somyaa
Basu, Samyadeep
Srinivasan, Soundararajan
Zhao, Wenlong
Chaudhari, Shreyas
Aghazadeh, Ehsan
Computation and Language
Retrieval Augmented Generation (RAG) is a widely used approach for leveraging external context in several natural language applications such as question answering and information retrieval. Yet, the exact nature in which a Language Model (LM) leverages this non-parametric memory or retrieved context isn't clearly understood. This paper mechanistically examines the RAG pipeline to highlight that LMs demonstrate a "shortcut'' effect and have a strong bias towards utilizing the retrieved context to answer questions, while relying minimally on model priors. We propose (a) Causal Mediation Analysis; for proving that parametric memory is minimally utilized when answering a question and (b) Attention Contributions and Knockouts for showing the last token residual stream do not get enriched from the subject token in the question, but gets enriched from tokens of RAG-context. We find this pronounced "shortcut'' behaviour to be true across both LLMs (e.g.,LlaMa) and SLMs (e.g., Phi)
title Quantifying reliance on external information over parametric knowledge during Retrieval Augmented Generation (RAG) using mechanistic analysis
topic Computation and Language
url https://arxiv.org/abs/2410.00857