Skip to content

R
RAGRights
Commit 9bd66690 authored by Delvallez Delvallez's avatar Delvallez Delvallez
Browse files
Options

Slides de soutenance de fin de stage

parent 8d286443
Hide whitespace changes
Inline Side-by-side
Showing with 1212 additions and 1 deletion
.gitignore
View file @ 9bd66690
......@@ -33,4 +33,5 @@ rapports_presentations/SoutenanceS9/SoutenanceXAI_S3_2026-01-07_10h.pdf
rapports_presentations/SoutenanceS9/brouillon.md
rapports_presentations/seminaireCA/indications_oral.md
rapports_presentations/seminaireCA/archives-utiles.md
rapports_presentations/rapportS10/reformatage-git.md
\ No newline at end of file
rapports_presentations/rapportS10/reformatage-git.md
rapports_presentations/SoutenanceS10/script.md
\ No newline at end of file
rapports_presentations/SoutenanceS10/main.tex 0 → 100644
View file @ 9bd66690
\documentclass[french, 12pt, aspectratio=2013]{beamer}
% valeur possibles : article, proc, book, report, letter, slides
% options possibles : twosides (recto-verso), 12pt, 11pt, ... (taille de police), draft (brouillon)
% Language setting
% Replace `english' with e.g. `spanish' to change the document language
\usepackage[french]{babel}
\usepackage[T1]{fontenc}
\usepackage{amssymb}
\usepackage{shadowtext}
% Useful packages
\usepackage{amsmath}
\usepackage{graphicx}
%tableau pleine largeur
\usepackage{tabularx}
%faire des dessins
\usepackage{tikz}
\usepackage{algorithm}
%\usepackage{fourier}
\usepackage{fontawesome}
\usepackage{color}
\newcommand{\md}[1]{\textcolor{cyan}{#1}}
\newcommand{\alh}[1]{\textcolor{teal}{#1}}
\newcommand{\append}{\texttt{append }}
\newcommand{\prepend}{\texttt{prepend }}
\newcommand{\replace}{\texttt{replace }}
\newcommand{\fullreplace}{\texttt{full\_replace }}
% Source - https://tex.stackexchange.com/a/680027
% Posted by C.F.G
% Retrieved 2026-08-27, License - CC BY-SA 4.0
\newenvironment<>{violetblock}[1]{%
\setbeamercolor{block title}{fg= violet!80!black,bg= violet!40!white}%
%\setbeamerfont{block title example}{family=\ssfamily}
\begin{block}{#1}{#2}}{\end{block}}
\usepackage[dvipsnames]{xcolor}
\usepackage{listings}
\usepackage{appendixnumberbeamer}
% Theme choice:
\usetheme{Boadilla}
\AtBeginSection[]{
\begin{frame}{Sommaire}
\tableofcontents[currentsection]
\end{frame}
}
\title{Interprétabilité Mécaniste pour l’Extraction d’Information dans les systèmes de RAG}
\author{Marine DELVALLEZ}
\setbeamertemplate{footline}[frame number]
\setbeamertemplate{navigation symbols}{}
\defbeamertemplate*{headline}{}
{%
\setbeamercolor{subsection in head/foot}{fg=white}
\begin{beamercolorbox}{section in head/foot}
\insertsectionnavigationhorizontal{\paperwidth}{}{}
\insertsubsectionnavigationhorizontal{\paperwidth}{}{}
\end{beamercolorbox}%
}
\begin{document}
\addtocounter{framenumber}{-1}
\begin{frame}[plain]
\includegraphics[scale=0.3]{images/Logo Minerve_RVB.jpg} \hfill
\includegraphics[scale=0.2]{images/France_2030_Logo_rouge_bleu_transparent.png} \hfill
\includegraphics[scale=0.3]{images/LIFO.png}
\centering
\Large Interprétabilité Mécaniste pour l’Extraction d’Information dans les systèmes de RAG\\
\large Soutenance de stage de Master 2\\
\vfill %\vspace{3mm}
\normalsize Marine DELVALLEZ \hfill
\textit{Master 2 Informatique GPEx Minerve}\\
\vfill %\vspace{3mm}
\small
\begin{tabular}{llll}
\emph{Jury} : & Thi-Bich-Hanh DAO &\emph{Tutrice de stage} : & Anaïs HALFTERMEYER\\
& Anaïs HALFTERMEYER & & \\
& Marie HÉNAULT
\end{tabular}
\vfill %\vspace{3mm}
\centering \footnotesize 2 Septembre 2026 \\
\end{frame}
\begin{frame}{Sommaire}
\tableofcontents[]
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{État de l'art}
\subsection{Retrieval Augmented Generation - RAG}
\begin{frame}{Retrieval Augmented Generation - RAG \cite{lewis_retrieval-augmented_2020}}
%\md{Définiton de RAG, Schéma de l'architecture de base, tâches et sous-tâche d'IR}
\includegraphics[width=\textwidth]{images/DefRAG.drawio.png}
\pause
\begin{columns}
\begin{column}{0.45\textwidth}
\begin{block}{Tâches du RAG}
\begin{itemize}
\item Question Answering
\item Fact Checking
\item Dialogue
\end{itemize}
\end{block}
\end{column}
\pause
\begin{column}{0.45\textwidth}
\begin{block}{Sous-tâches}
\begin{itemize}
\item \textbf<4>{Extraction de document}/ d'information
\item Synthèse
\item Génération de texte
\end{itemize}
\end{block}
\end{column}
\end{columns}
\end{frame}
\subsection{Explicabilité pour le RAG}
\begin{frame}{Explicabilité de l'IA et RAG}
%\md{XAI > XRAG > XIR : On retrouve les problèmes classiques du DL dans le RAG. Le RAG se divisant en deux composants, on distingue XIR, XGen et Mise en commun. On se concentre sur XIR}
\begin{alertblock}{}
On a besoin d'expliquer les systèmes RAG !
\end{alertblock}
{\footnotesize \cite{fan_survey_2024}}
\begin{picture}(370,170)
\put(0,65){\onslide<1>\includegraphics[width=\textwidth]{images/DefRAG.drawio.png}}
\put(0,65){\onslide<2>\includegraphics[width=\textwidth]{images/XRAG/XRAG_step1.drawio.png}}
\put(0,65){\onslide<3>\includegraphics[width=\textwidth]{images/XRAG/XRAG_step2.drawio.png}}
\put(0,0){\onslide<4>\includegraphics[width=\textwidth]{images/XRAG/XRAG_step3.drawio.png}}
\put(0,0){\onslide<5>\includegraphics[width=\textwidth]{images/XRAG/XRAG_step4.drawio.png}}
\end{picture}
\end{frame}
\begin{frame}{Explicabilité de l'IA et Interprétabilité Mécaniste \cite{somvanshi_bridging_2026}}
%\md{Définition et situer InterpMech, taxonomie de \cite{somvanshi_bridging_2026}}
\begin{center}
\includegraphics[width=0.65\textwidth]{images/InterprétabiliteMechaniste_Somvanshi-et-al_2026.drawio.png}
\end{center}
\pause
\begin{block}{Interprétabilité Mécaniste}
Explorer les \textbf{activations des composants} du modèle pour \textbf{identifier ceux qui provoquent ou influencent} un comportement cible
\end{block}
\end{frame}
\begin{frame}{Interprétabilité Mécaniste \cite{somvanshi_bridging_2026}}
%\md{Définition et situer InterpMech, taxonomie de \cite{somvanshi_bridging_2026}}
\begin{block}{Interprétabilité Mécaniste }
Explorer les \textbf{activations des composants} du modèle pour \textbf{identifier ceux qui provoquent ou influencent} un comportement cible
\end{block}
Plusieurs approches:
\begin{itemize}
\item Manual Circuit Tracing
\item Representation Analysis
\item Toy Models and Synthetic Tasks
\item \textbf<2>{Intervention-based Techniques}
\pause
\begin{itemize}
\item Activation Patching
\item ...
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Activation Patching pour l'IR \cite{chen_axiomatic_2024} : intuition }
%\md{intuition du fonctionnement de AP (animation utilisée au séminaire), évocation de MechIR comme implémentation}
\begin{picture}(350,150)
\put(0,0){\onslide<1>\includegraphics[width=\textwidth]{images/activationpatching/Step0_ActivationPatching.drawio.png}}
\put(0,0){\onslide<2>\includegraphics[width=\textwidth]{images/activationpatching/Step1_ActivationPatching.drawio.png}}
\put(0,0){\onslide<3>\includegraphics[width=\textwidth]{images/activationpatching/Step2_ActivationPatching.drawio.png}}
\put(0,0){\onslide<4>\includegraphics[width=\textwidth]{images/activationpatching/Step3_ActivationPatching.drawio.png}}
\put(0,0){\onslide<5>\includegraphics[width=\textwidth]{images/activationpatching/Step4_ActivationPatching.drawio.png}}
\put(0,0){\onslide<6->\includegraphics[width=\textwidth]{images/activationpatching/Step5_ActivationPatching.drawio.png}}
\end{picture}
%\onslide<7>{MechIR \cite{parry_mechir_2025} permet d'implémenter l'activation patching}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Objectifs de ce stage}
\begin{frame}{Hypothèses et Objectifs}
%\md{slide hypothèses et objectifs issue de présentation du séminaire + remis au goût du jour}
\begin{block}{Langue de spécialité \cite{cabre_terminologie_1998}}
Ensemble de sous-codes linguistiques spécifiques à un domaine de spécialité (lexique, syntaxe,...)
\end{block}
\begin{block}{Hypothèses}
\begin{description}
\item[H1 :] Le vocabulaire de spécialité (lexique) repose sur certains composants du modèle
\item[H2 :] L'activation patching permet de les localiser
\end{description}
\end{block}
\begin{block}{Objectifs}
\begin{itemize}
\item Appliquer l'activation patching à la localisation du vocabulaire de spécialité de la \textit{défense}
\item Proposer une méthodologie d'utilisation de MechIR
\end{itemize}
\end{block}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Préliminaires}
\begin{frame}{Cas d'application}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{images/RAGHN-perso.drawio.png}
\caption{Version simplifiée de RAG4HistoricalNewspapers \cite{tran_retrieval_2024}}
\end{figure}
\begin{columns}
\begin{column}{0.45\textwidth}
\begin{block}{mE5-small \cite{wang_multilingual_2024}}
\begin{itemize}
\item $\approx 17 M$ de paramètres
\item encodeur-only
\item bi-encodeur
\end{itemize}
\end{block}
\end{column}
\pause
\begin{column}{0.45\textwidth}
\begin{block}{MechIR \cite{parry_mechir_2025}}
\begin{itemize}
\item Alternative à TransformerLens \cite{nanda_transformerlens_2022} pour les encodeur-only
\item Implante des outils pour l'Activation Patching
\end{itemize}
\end{block}
\end{column}
\end{columns}
\end{frame}
\begin{frame}{Dataset}
%\md{Présentation du format et statistiques du dataset (1Q, 28 docs, 50\% pertinent, }
\begin{columns}
\begin{column}{0.6\textwidth}
\begin{block}{Contenu thématique}
\begin{itemize}
\item Extraits de documents issus des données du challenge RAG@EvalLLM
\item Question sur les \textit{limitations techniques des systèmes de drones}
\item Trois documents
\begin{itemize}
\item Limitations techniques du drone
\item Matériel de guerre - Code de la défense
\item Fausses informations pendant les JO/JP
\end{itemize}
\end{itemize}
\end{block}
\end{column}
\begin{column}{0.30\textwidth}
\begin{block}{Structure}
\begin{itemize}
\item 1 question
\item 28 extraits
\item 50\% de documents pertinents
\end{itemize}
\end{block}
\end{column}
\end{columns}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Expérimentations}
\subsection{Construction des perturbations}
\begin{frame}{Construire des perturbations}
\begin{violetblock}{Qu'est-ce qu'une bonne perturbation ?}
\begin{itemize}
\item Modification minimale du document \cite{chen_axiomatic_2024}
\item Modification significative de la sortie \cite{chen_axiomatic_2024}
\item Grande couverture du dataset
\end{itemize}
\end{violetblock}
\begin{block}{Perturbations proposées par \cite{parry_mechir_2025}}
\begin{description}
\item[\prepend] : ajouter un mot au début du document
\item[\append] : ajouter un mot à la fin du document
\item[\replace] : remplacer un mot par un autre
\end{description}
\end{block}
\pause
\underline{Exemple} : remplacer \textit{aérien} par \textit{poétique}
\begin{overlayarea}{13cm}{8mm}
\only<2> {\textit{\quad Certaines limitations sont plus particulièrement liées aux caractéristiques du vecteur aérien}}
\only<3-> {\textit{ \quad Certaines limitations sont plus particulièrement liées aux caractéristiques du vecteur \textbf{poétique}}}
\end{overlayarea}
\end{frame}
\begin{frame}{Choisir le mot à remplacer}
%\md{Les indicateurs proposés pour identifier les mots à remplacer}
\begin{columns}
\begin{column}{0.45\textwidth}
\begin{block}{Critères de choix des mots}
\begin{itemize}
\item Fréquence dans le dataset
\item TF-IDF du mot dans le dataset
\item Présence du mot dans la question
\item Porteur d'un sens marqué
\item \textbf{Appartenance à au moins un autre sous-code}
\end{itemize}
\end{block}
\end{column}
\pause
\begin{column}{0.5 \textwidth}
\fontsize{10}{0.8\baselineskip}\selectfont
\begin{tabular}{|l|l|l|}
\hline
Mot & Fréquence & TF-IDF \\ \hline \hline
\textbf{matériel} & 16 & 0.063 \\ \hline
\textbf{système} & 11 & 0.054 \\ \hline
\textbf{limitation} & 6 & 0.050 \\ \hline
\textbf{drone} & 10 & 0.044 \\ \hline
\textbf{défense} & 10 & 0.044 \\ \hline
prendre & 6 & 0.043 \\ \hline
compte & 5 & 0.040 \\ \hline
action & 8 & 0.040 \\ \hline
\textbf{article} & 9 & 0.040 \\ \hline
ministre & 8 & 0.039 \\ \hline
\textbf{aérien} & 7 & 0.037 \\ \hline
acteur & 6 & 0.034 \\ \hline
guerre & 8 & 0.034 \\ \hline
informationnel & 5 & 0.031 \\ \hline
information & 6 & 0.030 \\ \hline
\end{tabular}
\end{column}
\end{columns}
\end{frame}
\begin{frame}{Choisir le mot remplaçant}
%\md{Les 4 niveaux de perturbation, tableau du rapport revisité}
\fontsize{10}{0.8\baselineskip}\selectfont
Étant donné le mot cible (à remplacer)
\begin{tabular}{|c||p{0.30\linewidth}|p{0.25\linewidth}|p{0.25\linewidth}|}
\hline
Niv & Mot remplaçant & Phénomène cible & Exemple \\ \hline \hline
1 & Synonyme dans la langue de spécialité & Localisation du mot remplacé & aérien $\to$ de vol\\ \hline
2 & Synonyme dans le langage courant & Localisation du sens & aérien $\to$ poétique \\ \hline
\pause
3 & Mot sémantiquement incompatible & Localisation du sens (sans réparation) & aérien $\to$ épicurien \\ \hline
4 & Groupe de mots \textit{absurde} sémantiquement incompatibles & Localisation du sens (avec bruit) & aérien $\to$ acide tracté \\ \hline
\end{tabular}
\end{frame}
\begin{frame}{Perturbation de type \fullreplace}
%\md{définition, en quoi est-ce plus interessant (meilleure couverture du dataset)}
\begin{block}{Perturbation \fullreplace}
Appliquer une liste de perturbations \replace \textbf{simultanément}
\end{block}
\underline{Exemple} :
\begin{itemize}
\item remplacer \textit{aérien} par \textit{poétique}
\item remplacer \textit{limitation} par \textit{bornage}
\end{itemize}
\begin{overlayarea}{13cm}{8mm}
\only<1> {\textit{\quad Certaines limitations sont plus particulièrement liées aux caractéristiques du vecteur aérien}}
\only<2-> {\textit{ \quad Certaines \textbf{bornage}s sont plus particulièrement liées aux caractéristiques du vecteur \textbf{poétique}}}
\end{overlayarea}
\begin{block}{Buts}
\begin{itemize}
\item Augmenter la couverture du dataset
\item Effet d'ensemble
\end{itemize}
\end{block}
\end{frame}
\subsection{Évaluation d'une perturbation}
\begin{frame}{Distribution des scores de pertinence des documents}
%\md{évaluation proposée par \cite{parry_mechir_2025} : perturbation de la distribution des score de pertinence du document}
\begin{block}{}
Objectif : Distributions de scores différentes pour les documents et documents perturbés \cite{parry_mechir_2025}
\end{block}
\begin{center}
\includegraphics[width=0.9\textwidth]{images/DistribScoresPerturb_aerien-poetique.png}
\end{center}
\pause
\textbf{Interprétations} :
\begin{columns}
\begin{column}{0.35 \textwidth}
Perte de pertinence d'une partie des documents
\end{column}
\begin{column}{0.65 \textwidth}
Concentration d'une partie des documents autour d'un score\\
\faWarning \; Variation de représentation de la baseline
\end{column}
\end{columns}
\end{frame}
\begin{frame}{En observant l'impact de la perturbation sur la question}
\begin{block}{}
Objectif : Estimer l'impact d'une perturbation sur la question dans l'espace de représentation.
\end{block}
\begin{columns}
\begin{column}{0.6\textwidth}
\includegraphics[width = \textwidth]{images/Exp9QuestioncommeDocument.png}
\end{column}
\begin{column}{0.4\textwidth}
\fontsize{10}{0.8\baselineskip}\selectfont
\textit{Quelles sont les trois principales catégories de limitations techniques affectant les systèmes de drones, et comment chacune influence-t-elle l'efficacité opérationnelle de ces vecteurs ?}
\begin{block}{Mots ciblés}
\begin{columns}
\begin{column}{0.5\textwidth}
\begin{itemize}
\item drone
\item limitation
\end{itemize}
\end{column}
\begin{column}{0.5\textwidth}
\begin{itemize}
\item système
\item technique
\end{itemize}
\end{column}
\end{columns}
\end{block}
\end{column}
\end{columns}
\end{frame}
\begin{frame}{En séparant les documents pertinents et non-pertinents}
\begin{block}{}
Objectif : Identifier les documents perturbés et leur distribution en fonction de leur pertinence.
\end{block}
\includegraphics[width = \textwidth]{images/PerturbationScoreExemple_PosNeg.png}
$\Rightarrow$ Répartition des mots ciblés biaisée
\end{frame}
\begin{frame}{Confronter les perturbations entre elles}
\begin{block}{}
Objectif : Comparer l'impact des perturbations entre elles. Confirmer une progressivité. Choisir parmi plusieurs perturbations.
\end{block}
\includegraphics[width=\textwidth]{images/PerturbationScoreExemple_ParNiveaux.png}
\end{frame}
\subsection{Lecture des résultats de l'activation patching}
\begin{frame}{Matrice de sensibilité}
%\md{Exemple avec le graphique de \cite{parry_mechir_2025}. Ajoute des 4 autres graphiques + lecture accompagnée }
\centering
\begin{picture}(320,220)
\put(0,110){\includegraphics[width=0.4\textwidth]{images/MatriceSensibilite_new-mean_aerien-poetique.png}}
\put(160,110){\onslide<2->\includegraphics[width=0.4\textwidth]{images/MatriceSensibilite_new-std_aerien-poetique.png}}
\put(0,0){\onslide<3->\includegraphics[width=0.4\textwidth]{images/MatriceSensibilite_resc-mean_aerien-poetique.png}}
\put(160,0){\onslide<3->\includegraphics[width=0.4\textwidth]{images/MatriceSensibilite_resc-std_aerien-poetique.png}}
\end{picture}
\end{frame}
\begin{frame}{Identification des noeuds sensibles}
%\md{Construction du critère de sensibilité, matrice binaire pertinent ou non pour une perturbation (permet introduction progressive de la matrice ds noeuds sensibles)}
\fontsize{10}{0.8\baselineskip}\selectfont
\begin{violetblock} {Qu'est-ce qu'un noeud sensible?}
\begin{itemize}
\item Sensibilité au patching forte
\item ... plus forte que les autres noeuds
\item ... pour certains documents
\end{itemize}
\end{violetblock}
\begin{block}{Critère de sensibilité}
$\left(\left(\mid mean \mid > 0.1\right) \wedge \left(\mid resc\_mean \mid > 0.5\right) \right) \vee \left( \left(std > 0.1\right) \wedge \left(resc\_std > 0.5\right)\right)$
\end{block}
\centering
\begin{minipage}[c]{4.5cm}
\includegraphics[width=\textwidth]{images/MatriceSensibilite_new-mean_aerien-poetique.png}
\end{minipage}
$\Rightarrow$
\begin{minipage}[c]{4.5cm}
\includegraphics[width=0.8\textwidth]{images/MatNoeudsSensibles_aerien-poetique.png}
\end{minipage}
\end{frame}
\begin{frame}{Matrice des noeuds sensibles}
%\md{Visualisation de 2 matrices : perturbations ciblant un mot, perturbations d'un même niveau ~> pas toujours qqchose d’intéressant}
\includegraphics[width=0.49\textwidth]{images/NoeudsSensiblesAerien.png}
\includegraphics[width=0.49\textwidth]{images/NoeudsSensiblesNiv2.png}
\end{frame}
\subsection{Résultats}
\begin{frame}{Résultats}
\begin{block}{Quelques noeuds identifiés}
\begin{columns}
\begin{column}{0.5\textwidth}
\begin{tabular}{|c|p{0.55\linewidth}|}
\hline
Mot / Niveau & Noeuds identifiés \\ \hline
article & (11,4) \\ \hline
aérien & (11,6), (11,3), (10,9) \\ \hline
drone & (11,4) \\ \hline
défense & (11,2) \\ \hline
matériel & (11,4), (11,9) \\ \hline
\end{tabular}
\end{column}
\begin{column}{0.45\textwidth}
\begin{tabular}{|c|p{0.55\linewidth}|}
\hline
système & (11,1) \\ \hline
Niveau 3 & (11,3) \\ \hline
Niveau 4 & (1,7) \\ \hline
Full replace & (11,1), (11,5), (11,6), (11,3) \\ \hline
\end{tabular}
\end{column}
\end{columns}
\end{block}
\pause
\begin{block}{La méthodologie actuelle ...}
\begin{itemize}
\item ... met en évidence des noeuds sensibles \textit{pour une perturbation}
\item ... met en évidence des noeuds sensibles \textit{pour un mot perturbé} (parfois)
\item ... ne met pas en évidence de tendance \textit{à l'échelle du vocabulaire}
\end{itemize}
\end{block}
\end{frame}
\begin{frame}{Analyse}
\begin{violetblock}{\textbf{H1 :} Le vocabulaire de spécialité (lexique) repose sur certains composants du modèle}
\begin{itemize}
\item On peut identifier une spécialisation de certains composants à l'échelle du mot
\item Sans savoir si elle relève du vocabulaire de spécialité, du mot ou du bruit
\end{itemize}
$\Rightarrow$ On ne peut pas conclure pour l'instant
\end{violetblock}
\begin{violetblock}{\textbf{H2 :} L'activation patching permet de les localiser}
\begin{itemize}
\item L'activation patching permet de localiser les composants sensibles d'un modèle
\end{itemize}
$\Rightarrow$ On ne peut pas conclure
\end{violetblock}
%\md{reprendre les hypothèses et conclure dessus}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Conclusion et perspectives}
\begin{frame}{À retenir}
\begin{block}{Bonnes pratiques}
\begin{itemize}
\item Perturbations couvrantes
\item Perturbations pertinentes (cas de \textit{article})
\item Vigilance et prise en compte des biais
\item Vigilance pour les noeuds de la dernière couche
\end{itemize}
\end{block}
\begin{block}{Résultats}
\begin{itemize}
\item L'activation patching permet de localiser les composants sensibles au sens de certains mots
\end{itemize}
\end{block}
\end{frame}
\begin{frame}{Perspectives d'amélioration}
\begin{itemize}[<+->]
\item Fonder et outiller l'échelle de niveau des perturbations \cite{greimas_semiotique_1979}
\item Méthodologie pour mettre en perspective différents datasets
\item Diversification des perturbations (\prepend, \append, ...)
\item Étendre à la langue de spécialité
\end{itemize}
\end{frame}
\begin{frame}{Intégration dans l'XRAG}
%\md{lien avec le projet de thèse, projet dans lequel s'intègre le travail du stage, perspectives pour aller au bout de se projet}
\begin{itemize}[<+->]
\item Amélioration de la méthodologie
\item Transformer les noeuds identifiés en cartographie
\item Protocole d'extraction d'explication
\item Rendre les explications accessibles à l'utilisateur non-informaticien
\vspace{2mm}
\item Protocole d'extraction d'améliorations
\end{itemize}
\end{frame}
\begin{frame}{Conclusion sur les travaux du stage}
%\md{liste à puce récap de ce qui a été présenté : Def de RAG, InterpMEch et AP + modèle et dataset challenge + Méthode de construction de peturbation + outils de visualistaion des résultats de AP}
\begin{itemize}
\item Défrichage bibliographique
\begin{itemize}
\item RAG et RA-LLM
\item Explicabilité pour le RAG et le RA-LLM
\item Interprétabilité mécaniste
\item Activation Patching
\end{itemize}
\item Challenge RAG@EvalLLM
\begin{itemize}
\item Reprise d'un système RAG
\item Structuration itérative d'un dataset à la main
\end{itemize}
\item Proposition d'une méthode d'utilisation de MechIR
\item Proposition d'une méthode d'interprétation de l'activation patching
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\appendix
\bibliographystyle{alpha} %apalike
\begin{frame}[allowframebreaks]{Références}
\bibliography{sample.bib}
\end{frame}
\section{Annexes}
\begin{frame}{Activation Patching pour l'IR \cite{chen_axiomatic_2024}}
\fontsize{10}{0.8\baselineskip}\selectfont % pour écrire plus petit en gardant l'interligne standard de beamer
Soit $Q \times D \subset \mathcal{Q}\times\mathcal{D}$ l'ensemble des paires question, document. Soit $Q \times \tilde{D}$ le même ensemble mais avec les documents perturbés à la place
\begin{enumerate}
\item Faire une passe pour toutes les paires $Q\times D$
\begin{itemize}
\item noter $o_{n}^e$ la sortie de chaque nœud $n, \forall e \in Q\times D$
\item noter $p_D$ la performance du modèle sur les données non perturbées
\end{itemize}
\item Faire une passe pour toutes les paires $Q\times \tilde{D}$
\begin{itemize}
\item noter $o_{n}^{\tilde{e}}$ la sortie de chaque nœud $n, \forall \tilde{e} \in Q\times \tilde{D}$
\item noter $p_{\tilde{D}}$ la performance du modèle sur les données perturbées
\end{itemize}
\item On renomme $D, e, \tilde{D}$ et $\tilde{e}$ respectivement
\begin{itemize}
\item $\hat{D}, \hat{e}, \check{D}$ et $\check{e}$ si $p_D > p_{\tilde{D}}$
\item $\check{D}, \check{e}, \hat{D}$ et $\hat{e}$ sinon
\end{itemize}
Pour chaque nœud $n$
\item Faire une passe de $Q\times\check{D}$ en remplaçant $o_{i,j}^{\check{e}}$ par $o_{n}^{\hat{e}}$ pour chaque $\check{e}$. On note la performance $\bar{p}_n$
\item $P_{n} = \frac{\bar{p}_n - p_{\hat{D}} }{p_{\check{D}} - p_{\hat{D}}}$ exprime l'impact de la perturbation sur la performance du modèle pour le nœud $n$
\end{enumerate}
\end{frame}
\begin{frame}{Liste des perturbations pratiquées}
\begin{tabular}{|l||l|l|l|l|}
\hline
Mot remplacé & Niveau 1 & Niveau 2 & Niveau 3 & Niveau 4 \\ \hline \hline
aérien & de vol & poétique & épicurien & acide tracté \\ \hline
article & sujet & déterminant & table & luge pontificale \\ \hline
défense & protection & corne & fromage & poire retournée \\ \hline
drone & vecteur & bourdon & gazon & ver de verre \\ \hline
limitation & contrainte & bornage & animation & plastique anarchique \\ \hline
matériel & outil & concret & post-it & porte sans poivre \\ \hline
système & dispositif & ensemble & panneau & esperluette en vent \\ \hline
technique & spécialisé & difficile & magique & feuille entubée \\ \hline
\end{tabular}
\end{frame}
\begin{frame}[plain]{}
\centering
\includegraphics[width=0.8\textwidth]{images/MatriceNoeudsSensibles_all.png}
\end{frame}
\end{document}
rapports_presentations/SoutenanceS10/sample.bib 0 → 100644
View file @ 9bd66690
@misc{nanda_transformerlens_2022,
title = {{TransformerLens}},
url = {https://github.com/TransformerLensOrg/TransformerLens},
author = {Nanda, Neel and Bloom, Joseph},
year = {2022},
}
@misc{wolf_huggingfaces_2020,
title = {{HuggingFace}'s {Transformers}: {State}-of-the-art {Natural} {Language} {Processing}},
shorttitle = {{HuggingFace}'s {Transformers}},
url = {http://arxiv.org/abs/1910.03771},
doi = {10.48550/arXiv.1910.03771},
abstract = {Recent progress in natural language processing has been driven by advances in both model architecture and model pretraining. Transformer architectures have facilitated building higher-capacity models and pretraining has made it possible to effectively utilize this capacity for a wide variety of tasks. {\textbackslash}textit\{Transformers\} is an open-source library with the goal of opening up these advances to the wider machine learning community. The library consists of carefully engineered state-of-the art Transformer architectures under a unified API. Backing this library is a curated collection of pretrained models made by and available for the community. {\textbackslash}textit\{Transformers\} is designed to be extensible by researchers, simple for practitioners, and fast and robust in industrial deployments. The library is available at {\textbackslash}url\{https://github.com/huggingface/transformers\}.},
urldate = {2026-08-25},
publisher = {arXiv},
author = {Wolf, Thomas and Debut, Lysandre and Sanh, Victor and Chaumond, Julien and Delangue, Clement and Moi, Anthony and Cistac, Pierric and Rault, Tim and Louf, Rémi and Funtowicz, Morgan and Davison, Joe and Shleifer, Sam and Platen, Patrick von and Ma, Clara and Jernite, Yacine and Plu, Julien and Xu, Canwen and Scao, Teven Le and Gugger, Sylvain and Drame, Mariama and Lhoest, Quentin and Rush, Alexander M.},
month = jul,
year = {2020},
note = {arXiv:1910.03771 [cs.CL]},
keywords = {Computer Science - Computation and Language},
}
@incollection{paszke_pytorch_2019,
address = {Red Hook, NY, USA},
title = {{PyTorch}: an imperative style, high-performance deep learning library},
abstract = {Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs.In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks.},
booktitle = {Proceedings of the 33rd {International} {Conference} on {Neural} {Information} {Processing} {Systems}},
publisher = {Curran Associates Inc.},
author = {Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and Köpf, Andreas and Yang, Edward and DeVito, Zach and Raison, Martin and Tejani, Alykhan and Chilamkurthy, Sasank and Steiner, Benoit and Fang, Lu and Bai, Junjie and Chintala, Soumith},
year = {2019},
}
@article{majewska_semantic_2021,
title = {Semantic {Data} {Set} {Construction} from {Human} {Clustering} and {Spatial} {Arrangement}},
volume = {47},
issn = {0891-2017},
url = {https://doi.org/10.1162/coli_a_00396},
doi = {10.1162/coli_a_00396},
abstract = {Research into representation learning models of lexical semantics usually utilizes some form of intrinsic evaluation to ensure that the learned representations reflect human semantic judgments. Lexical semantic similarity estimation is a widely used evaluation method, but efforts have typically focused on pairwise judgments of words in isolation, or are limited to specific contexts and lexical stimuli. There are limitations with these approaches that either do not provide any context for judgments, and thereby ignore ambiguity, or provide very specific sentential contexts that cannot then be used to generate a larger lexical resource. Furthermore, similarity between more than two items is not considered. We provide a full description and analysis of our recently proposed methodology for large-scale data set construction that produces a semantic classification of a large sample of verbs in the first phase, as well as multi-way similarity judgments made within the resultant semantic classes in the second phase. The methodology uses a spatial multi-arrangement approach proposed in the field of cognitive neuroscience for capturing multi-way similarity judgments of visual stimuli. We have adapted this method to handle polysemous linguistic stimuli and much larger samples than previous work. We specifically target verbs, but the method can equally be applied to other parts of speech. We perform cluster analysis on the data from the first phase and demonstrate how this might be useful in the construction of a comprehensive verb resource. We also analyze the semantic information captured by the second phase and discuss the potential of the spatially induced similarity judgments to better reflect human notions of word similarity. We demonstrate how the resultant data set can be used for fine-grained analyses and evaluation of representation learning models on the intrinsic tasks of semantic clustering and semantic similarity. In particular, we find that stronger static word embedding methods still outperform lexical representations emerging from more recent pre-training methods, both on word-level similarity and clustering. Moreover, thanks to the data set’s vast coverage, we are able to compare the benefits of specializing vector representations for a particular type of external knowledge by evaluating FrameNet- and VerbNet-retrofitted models on specific semantic domains such as “Heat” or “Motion.”},
number = {1},
journal = {Computational Linguistics},
author = {Majewska, Olga and McCarthy, Diana and van den Bosch, Jasper J. F. and Kriegeskorte, Nikolaus and Vulić, Ivan and Korhonen, Anna},
month = apr,
year = {2021},
pages = {69--116},
}
@book{greimas_semiotique_1979,
address = {Paris},
title = {Sémiotique. {Dictionnaire} raisonné de la théorie du langage},
publisher = {Hachette},
author = {Greimas, Algirdas Julien and Courtès, Joseph},
year = {1979},
}
@misc{cohere_rerank_2024,
title = {Rerank {Multilingual} v3.0},
url = {https://docs.cohere.com/docs/rerank},
urldate = {2026-08-24},
publisher = {Cohere},
author = {{Cohere}},
year = {2024},
}
@book{cabre_terminologie_1998,
series = {U. {Linguistique}},
title = {La terminologie : théorie, méthode et application / {Maria} {Teresa} {Cabré}},
isbn = {9782200218911},
shorttitle = {La terminologie},
url = {https://www.bm-vitry94.fr/Default/doc/SYRACUSE/603932/la-terminologie-theorie-methode-et-application-maria-teresa-cabre},
abstract = {Ce manuel passe en revue l'histoire de la terminologie, des origines à nos jours, et analyse ses fondements théoriques et méthodologiques par rapport aux disciplines connexes.{\textless}br/{\textgreater}{\textless}br/{\textgreater}},
language = {français},
urldate = {2026-08-20},
publisher = {Armand Colin. Paris},
author = {Cabré, Maria Teresa},
year = {1998},
keywords = {Lexicologie, Locutions françaises, Terminologie},
}
@inproceedings{wang_query2doc_2023,
address = {Singapore},
title = {Query2doc: {Query} {Expansion} with {Large} {Language} {Models}},
shorttitle = {Query2doc},
url = {https://aclanthology.org/2023.emnlp-main.585/},
doi = {10.18653/v1/2023.emnlp-main.585},
abstract = {This paper introduces a simple yet effective query expansion approach, denoted as query2doc, to improve both sparse and dense retrieval systems. The proposed method first generates pseudo-documents by few-shot prompting large language models (LLMs), and then expands the query with generated pseudo documents. LLMs are trained on web-scale text corpora and are adept at knowledge memorization. The pseudo-documents from LLMs often contain highly relevant information that can aid in query disambiguation and guide the retrievers. Experimental results demonstrate that query2doc boosts the performance of BM25 by 3\% to 15\% on ad-hoc IR datasets, such as MS-MARCO and TREC DL, without any model fine-tuning. Furthermore, our method also benefits state-of-the-art dense retrievers in terms of both in-domain and out-of-domain results.},
urldate = {2026-08-20},
booktitle = {Proceedings of the 2023 {Conference} on {Empirical} {Methods} in {Natural} {Language} {Processing}},
publisher = {Association for Computational Linguistics},
author = {Wang, Liang and Yang, Nan and Wei, Furu},
editor = {Bouamor, Houda and Pino, Juan and Bali, Kalika},
month = dec,
year = {2023},
pages = {9414--9423},
}
@article{sparck_jones_statistical_1972,
title = {A {STATISTICAL} {INTERPRETATION} {OF} {TERM} {SPECIFICITY} {AND} {ITS} {APPLICATION} {IN} {RETRIEVAL}},
volume = {28},
issn = {0022-0418},
url = {http://www.emerald.com/jd/article/28/1/11-21/196227},
doi = {10.1108/eb026526},
language = {en},
number = {1},
urldate = {2026-08-20},
journal = {Journal of Documentation},
author = {Sparck Jones, Karen},
month = jan,
year = {1972},
pages = {11--21},
}
@misc{ma_query_2023,
title = {Query {Rewriting} for {Retrieval}-{Augmented} {Large} {Language} {Models}},
url = {http://arxiv.org/abs/2305.14283},
doi = {10.48550/arXiv.2305.14283},
abstract = {Large Language Models (LLMs) play powerful, black-box readers in the retrieve-then-read pipeline, making remarkable progress in knowledge-intensive tasks. This work introduces a new framework, Rewrite-Retrieve-Read instead of the previous retrieve-then-read for the retrieval-augmented LLMs from the perspective of the query rewriting. Unlike prior studies focusing on adapting either the retriever or the reader, our approach pays attention to the adaptation of the search query itself, for there is inevitably a gap between the input text and the needed knowledge in retrieval. We first prompt an LLM to generate the query, then use a web search engine to retrieve contexts. Furthermore, to better align the query to the frozen modules, we propose a trainable scheme for our pipeline. A small language model is adopted as a trainable rewriter to cater to the black-box LLM reader. The rewriter is trained using the feedback of the LLM reader by reinforcement learning. Evaluation is conducted on downstream tasks, open-domain QA and multiple-choice QA. Experiments results show consistent performance improvement, indicating that our framework is proven effective and scalable, and brings a new framework for retrieval-augmented LLM.},
urldate = {2026-08-19},
publisher = {arXiv},
author = {Ma, Xinbei and Gong, Yeyun and He, Pengcheng and Zhao, Hai and Duan, Nan},
month = oct,
year = {2023},
note = {arXiv:2305.14283 [cs.CL]},
keywords = {Computer Science - Computation and Language},
}
@article{huang_survey_2025,
title = {A {Survey} on {Hallucination} in {Large} {Language} {Models}: {Principles}, {Taxonomy}, {Challenges}, and {Open} {Questions}},
volume = {43},
issn = {1046-8188, 1558-2868},
shorttitle = {A {Survey} on {Hallucination} in {Large} {Language} {Models}},
url = {https://dl.acm.org/doi/10.1145/3703155},
doi = {10.1145/3703155},
abstract = {The emergence of large language models (LLMs) has marked a significant breakthrough in natural language processing (NLP), fueling a paradigm shift in information acquisition. Nevertheless, LLMs are prone to hallucination, generating plausible yet nonfactual content. This phenomenon raises significant concerns over the reliability of LLMs in real-world information retrieval (IR) systems and has attracted intensive research to detect and mitigate such hallucinations. Given the open-ended general-purpose attributes inherent to LLMs, LLM hallucinations present distinct challenges that diverge from prior task-specific models. This divergence highlights the urgency for a nuanced understanding and comprehensive overview of recent advances in LLM hallucinations. In this survey, we begin with an innovative taxonomy of hallucination in the era of LLM and then delve into the factors contributing to hallucinations. Subsequently, we present a thorough overview of hallucination detection methods and benchmarks. Our discussion then transfers to representative methodologies for mitigating LLM hallucinations. Additionally, we delve into the current limitations faced by retrieval-augmented LLMs in combating hallucinations, offering insights for developing more robust IR systems. Finally, we highlight the promising research directions on LLM hallucinations, including hallucination in large vision-language models and understanding of knowledge boundaries in LLM hallucinations.},
language = {en},
number = {2},
urldate = {2026-08-19},
journal = {ACM Transactions on Information Systems},
author = {Huang, Lei and Yu, Weijiang and Ma, Weitao and Zhong, Weihong and Feng, Zhangyin and Wang, Haotian and Chen, Qianglong and Peng, Weihua and Feng, Xiaocheng and Qin, Bing and Liu, Ting},
month = mar,
year = {2025},
pages = {1--55},
}
@misc{artifex_software_pymupdf4llm_2026,
title = {{PyMuPDF4LLM}: {PDF} {Markdown} {Extraction} for {LLM} workflows},
url = {https://pymupdf.readthedocs.io/en/latest/pymupdf4llm/},
author = {{Artifex Software}},
year = {2026},
}
@inproceedings{chen_seeing_2025,
address = {Abu Dhabi, UAE},
title = {Seeing {Beyond}: {Enhancing} {Visual} {Question} {Answering} with {Multi}-{Modal} {Retrieval}},
shorttitle = {Seeing {Beyond}},
url = {https://aclanthology.org/2025.coling-industry.35/},
abstract = {Multi-modal Large language models (MLLMs) have made significant strides in complex content understanding and reasoning. However, they still suffer from model hallucination and lack of specific knowledge when facing challenging questions. To address these limitations, retrieval augmented generation (RAG) has emerged as an effective solution. While incorporating knowledge has led to improvements, it also highlights the need for a more robust knowledge selection strategy. For multi-modal tasks, such as visual question answering (VQA), integrating all modalities is crucial in providing comprehensive information for accurate answers. Therefore, we propose to construct an encoder model for extracting joint embedding from all modalities, enabling alignment between the corresponding query and knowledge through contrastive learning. To further improve performance, we introduce an additional MLLM re-selection step, which selects the best matching knowledge from the top-k retrieved results of our alignment model. We evaluated our method, SeBe-VQA, on the Encyclopedic VQA dataset. Our knowledge retrieval results demonstrate the benefit of our multi-modal framework. By incorporating the retrieved knowledge along with the question, we achieve a significant performance improvement compared with the previous method and scenarios without knowledge provision.},
urldate = {2026-08-19},
booktitle = {Proceedings of the 31st {International} {Conference} on {Computational} {Linguistics}: {Industry} {Track}},
publisher = {Association for Computational Linguistics},
author = {Chen, Boqi and Khare, Anuj and Kumar, Gaurav and Akula, Arjun and Narayana, Pradyumna},
editor = {Rambow, Owen and Wanner, Leo and Apidianaki, Marianna and Al-Khalifa, Hend and Eugenio, Barbara Di and Schockaert, Steven and Darwish, Kareem and Agarwal, Apoorv},
month = jan,
year = {2025},
pages = {410--421},
}
@inproceedings{macavaney_simplified_2021,
title = {Simplified {Data} {Wrangling} with ir\_datasets},
booktitle = {{SIGIR}},
author = {MacAvaney, Sean and Yates, Andrew and Feldman, Sergey and Downey, Doug and Cohan, Arman and Goharian, Nazli},
year = {2021},
}
@misc{hofstatter_efficiently_2021,
title = {Efficiently {Teaching} an {Effective} {Dense} {Retriever} with {Balanced} {Topic} {Aware} {Sampling}},
url = {http://arxiv.org/abs/2104.06967},
doi = {10.48550/arXiv.2104.06967},
abstract = {A vital step towards the widespread adoption of neural retrieval models is their resource efficiency throughout the training, indexing and query workflows. The neural IR community made great advancements in training effective dual-encoder dense retrieval (DR) models recently. A dense text retrieval model uses a single vector representation per query and passage to score a match, which enables low-latency first stage retrieval with a nearest neighbor search. Increasingly common, training approaches require enormous compute power, as they either conduct negative passage sampling out of a continuously updating refreshing index or require very large batch sizes for in-batch negative sampling. Instead of relying on more compute capability, we introduce an efficient topic-aware query and balanced margin sampling technique, called TAS-Balanced. We cluster queries once before training and sample queries out of a cluster per batch. We train our lightweight 6-layer DR model with a novel dual-teacher supervision that combines pairwise and in-batch negative teachers. Our method is trainable on a single consumer-grade GPU in under 48 hours (as opposed to a common configuration of 8x V100s). We show that our TAS-Balanced training method achieves state-of-the-art low-latency (64ms per query) results on two TREC Deep Learning Track query sets. Evaluated on NDCG@10, we outperform BM25 by 44\%, a plainly trained DR by 19\%, docT5query by 11\%, and the previous best DR model by 5\%. Additionally, TAS-Balanced produces the first dense retriever that outperforms every other method on recall at any cutoff on TREC-DL and allows more resource intensive re-ranking models to operate on fewer passages to improve results further.},
urldate = {2026-08-18},
publisher = {arXiv},
author = {Hofstätter, Sebastian and Lin, Sheng-Chieh and Yang, Jheng-Hong and Lin, Jimmy and Hanbury, Allan},
month = may,
year = {2021},
note = {arXiv:2104.06967 [cs.IR]},
keywords = {Computer Science - Computation and Language, Computer Science - Information Retrieval},
}
@misc{geiger_causal_2021,
title = {Causal {Abstractions} of {Neural} {Networks}},
url = {http://arxiv.org/abs/2106.02997},
doi = {10.48550/arXiv.2106.02997},
abstract = {Structural analysis methods (e.g., probing and feature attribution) are increasingly important tools for neural network analysis. We propose a new structural analysis method grounded in a formal theory of causal abstraction that provides rich characterizations of model-internal representations and their roles in input/output behavior. In this method, neural representations are aligned with variables in interpretable causal models, and then interchange interventions are used to experimentally verify that the neural representations have the causal properties of their aligned variables. We apply this method in a case study to analyze neural models trained on Multiply Quantified Natural Language Inference (MQNLI) corpus, a highly complex NLI dataset that was constructed with a tree-structured natural logic causal model. We discover that a BERT-based model with state-of-the-art performance successfully realizes parts of the natural logic model's causal structure, whereas a simpler baseline model fails to show any such structure, demonstrating that BERT representations encode the compositional structure of MQNLI.},
urldate = {2026-08-18},
publisher = {arXiv},
author = {Geiger, Atticus and Lu, Hanson and Icard, Thomas and Potts, Christopher},
month = oct,
year = {2021},
note = {arXiv:2106.02997 [cs.AI]},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Machine Learning},
}
@misc{wang_multilingual_2024,
title = {Multilingual {E5} {Text} {Embeddings}: {A} {Technical} {Report}},
shorttitle = {Multilingual {E5} {Text} {Embeddings}},
url = {https://arxiv.org/abs/2402.05672v1},
abstract = {This technical report presents the training methodology and evaluation results of the open-source multilingual E5 text embedding models, released in mid-2023. Three embedding models of different sizes (small / base / large) are provided, offering a balance between the inference efficiency and embedding quality. The training procedure adheres to the English E5 model recipe, involving contrastive pre-training on 1 billion multilingual text pairs, followed by fine-tuning on a combination of labeled datasets. Additionally, we introduce a new instruction-tuned embedding model, whose performance is on par with state-of-the-art, English-only models of similar sizes. Information regarding the model release can be found at https://github.com/microsoft/unilm/tree/master/e5 .},
language = {en},
urldate = {2026-06-03},
author = {Wang, Liang and Yang, Nan and Huang, Xiaolong and Yang, Linjun and Majumder, Rangan and Wei, Furu},
month = feb,
year = {2024},
}
@inproceedings{wang_minilmv2_2021,
address = {Online},
title = {{MiniLMv2}: {Multi}-{Head} {Self}-{Attention} {Relation} {Distillation} for {Compressing} {Pretrained} {Transformers}},
shorttitle = {{MiniLMv2}},
url = {https://aclanthology.org/2021.findings-acl.188/},
doi = {10.18653/v1/2021.findings-acl.188},
urldate = {2026-06-02},
booktitle = {Findings of the {Association} for {Computational} {Linguistics}: {ACL}-{IJCNLP} 2021},
publisher = {Association for Computational Linguistics},
author = {Wang, Wenhui and Bao, Hangbo and Huang, Shaohan and Dong, Li and Wei, Furu},
editor = {Zong, Chengqing and Xia, Fei and Li, Wenjie and Navigli, Roberto},
month = aug,
year = {2021},
pages = {2140--2151},
}
@article{somvanshi_bridging_2026,
title = {Bridging the {Black} {Box}: {A} {Survey} on {Mechanistic} {Interpretability} in {AI}},
volume = {58},
issn = {0360-0300, 1557-7341},
shorttitle = {Bridging the {Black} {Box}},
url = {https://dl.acm.org/doi/10.1145/3787104},
doi = {10.1145/3787104},
abstract = {Mechanistic interpretability seeks to reverse-engineer the internal logic of neural networks by uncovering human-understandable circuits, algorithms, and causal structures that drive model behavior. Unlike post hoc explanations that describe what models do, this paradigm focuses on why and how they compute, tracing information flow through neurons, attention heads, and activation pathways. This survey provides a high-level synthesis of the field-highlighting its motivation, conceptual foundations, and methodological taxonomy rather than enumerating individual techniques. We organize mechanistic interpretability across three abstraction layers—
neurons
,
circuits
, and
algorithms
—and three evaluation perspectives:
behavioral
,
counterfactual
, and
causal
. We further discuss representative approaches and toolchains that enable structural analysis of modern AI systems, outlining how mechanistic interpretability bridges theoretical insights with practical transparency. Despite rapid progress, challenges persist in scaling these analyses to frontier models, resolving polysemantic representations, and establishing standardized causal benchmarks. By connecting historical evolution, current methodologies, and emerging research directions, this survey aims to provide an integrative framework for understanding how mechanistic interpretability can support transparency, reliability, and governance in large-scale AI.},
language = {en},
number = {8},
urldate = {2026-06-01},
journal = {ACM Computing Surveys},
author = {Somvanshi, Shriyank and Islam, Md Monzurul and Rafe, Amir and Tusti, Anannya Ghosh and Chakraborty, Arka and Baitullah, Anika and Chowdhury, Tausif Islam and Alnawmasi, Nawaf and Dutta, Anandi and Das, Subasish},
month = jun,
year = {2026},
pages = {1--35},
}
@inproceedings{chen_axiomatic_2024,
address = {Washington DC USA},
title = {Axiomatic {Causal} {Interventions} for {Reverse} {Engineering} {Relevance} {Computation} in {Neural} {Retrieval} {Models}},
isbn = {9798400704314},
url = {https://dl.acm.org/doi/10.1145/3626772.3657841},
doi = {10.1145/3626772.3657841},
language = {en},
urldate = {2026-05-21},
booktitle = {Proceedings of the 47th {International} {ACM} {SIGIR} {Conference} on {Research} and {Development} in {Information} {Retrieval}},
publisher = {ACM},
author = {Chen, Catherine and Merullo, Jack and Eickhoff, Carsten},
month = jul,
year = {2024},
pages = {1401--1410},
}
@inproceedings{tran_retrieval_2024,
address = {Hong Kong China},
title = {Retrieval {Augmented} {Generation} for {Historical} {Newspapers}},
isbn = {9798400710933},
url = {https://dl.acm.org/doi/10.1145/3677389.3702542},
doi = {10.1145/3677389.3702542},
language = {en},
urldate = {2026-05-21},
booktitle = {Proceedings of the 24th {ACM}/{IEEE} {Joint} {Conference} on {Digital} {Libraries}},
publisher = {ACM},
author = {Tran, The Trung and González-Gallardo, Carlos-Emiliano and Doucet, Antoine},
month = dec,
year = {2024},
pages = {1--5},
}
@misc{wang_text_2022,
title = {Text {Embeddings} by {Weakly}-{Supervised} {Contrastive} {Pre}-training},
url = {https://arxiv.org/abs/2212.03533v2},
abstract = {This paper presents E5, a family of state-of-the-art text embeddings that transfer well to a wide range of tasks. The model is trained in a contrastive manner with weak supervision signals from our curated large-scale text pair dataset (called CCPairs). E5 can be readily used as a general-purpose embedding model for any tasks requiring a single-vector representation of texts such as retrieval, clustering, and classification, achieving strong performance in both zero-shot and fine-tuned settings. We conduct extensive evaluations on 56 datasets from the BEIR and MTEB benchmarks. For zero-shot settings, E5 is the first model that outperforms the strong BM25 baseline on the BEIR retrieval benchmark without using any labeled data. When fine-tuned, E5 obtains the best results on the MTEB benchmark, beating existing embedding models with 40x more parameters.},
language = {en},
urldate = {2026-05-20},
author = {Wang, Liang and Yang, Nan and Huang, Xiaolong and Jiao, Binxing and Yang, Linjun and Jiang, Daxin and Majumder, Rangan and Wei, Furu},
month = dec,
year = {2022},
}
@misc{lu_bm25s_2024,
title = {{BM25S}: {Orders} of magnitude faster lexical search via eager sparse scoring},
shorttitle = {{BM25S}},
url = {http://arxiv.org/abs/2407.03618},
doi = {10.48550/arXiv.2407.03618},
abstract = {We introduce BM25S, an efficient Python-based implementation of BM25 that only depends on Numpy and Scipy. BM25S achieves up to a 500x speedup compared to the most popular Python-based framework by eagerly computing BM25 scores during indexing and storing them into sparse matrices. It also achieves considerable speedups compared to highly optimized Java-based implementations, which are used by popular commercial products. Finally, BM25S reproduces the exact implementation of five BM25 variants based on Kamphuis et al. (2020) by extending eager scoring to non-sparse variants using a novel score shifting method. The code can be found at https://github.com/xhluca/bm25s},
urldate = {2026-03-31},
publisher = {arXiv},
author = {Lù, Xing Han},
month = jul,
year = {2024},
note = {arXiv:2407.03618},
keywords = {Computer Science - Computation and Language, Computer Science - Information Retrieval},
}
@inproceedings{parry_mechir_2025,
series = {Lecture {Notes} in {Computer} {Science}},
title = {{MechIR}: {A} {Mechanistic} {Interpretability} {Framework} for {Information} {Retrieval}},
volume = {15576},
url = {https://doi.org/10.1007/978-3-031-88720-8\_16},
doi = {10.1007/978-3-031-88720-8_16},
booktitle = {Advances in {Information} {Retrieval} - 47th {European} {Conference} on {Information} {Retrieval}, {ECIR} 2025, {Lucca}, {Italy}, {April} 6-10, 2025, {Proceedings}, {Part} {V}},
publisher = {Springer},
author = {Parry, Andrew and Chen, Catherine and Eickhoff, Carsten and MacAvaney, Sean},
year = {2025},
pages = {89--95},
}
@misc{izacard_unsupervised_2022,
title = {Unsupervised {Dense} {Information} {Retrieval} with {Contrastive} {Learning}},
url = {http://arxiv.org/abs/2112.09118},
doi = {10.48550/arXiv.2112.09118},
abstract = {Recently, information retrieval has seen the emergence of dense retrievers, using neural networks, as an alternative to classical sparse methods based on term-frequency. These models have obtained state-of-the-art results on datasets and tasks where large training sets are available. However, they do not transfer well to new applications with no training data, and are outperformed by unsupervised term-frequency methods such as BM25. In this work, we explore the limits of contrastive learning as a way to train unsupervised dense retrievers and show that it leads to strong performance in various retrieval settings. On the BEIR benchmark our unsupervised model outperforms BM25 on 11 out of 15 datasets for the Recall@100. When used as pre-training before fine-tuning, either on a few thousands in-domain examples or on the large MS{\textasciitilde}MARCO dataset, our contrastive model leads to improvements on the BEIR benchmark. Finally, we evaluate our approach for multi-lingual retrieval, where training data is even scarcer than for English, and show that our approach leads to strong unsupervised performance. Our model also exhibits strong cross-lingual transfer when fine-tuned on supervised English data only and evaluated on low resources language such as Swahili. We show that our unsupervised models can perform cross-lingual retrieval between different scripts, such as retrieving English documents from Arabic queries, which would not be possible with term matching methods.},
urldate = {2026-03-16},
publisher = {arXiv},
author = {Izacard, Gautier and Caron, Mathilde and Hosseini, Lucas and Riedel, Sebastian and Bojanowski, Piotr and Joulin, Armand and Grave, Edouard},
month = aug,
year = {2022},
note = {arXiv:2112.09118},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language, Computer Science - Information Retrieval},
}
@article{izacard_atlas_2023,
title = {Atlas: few-shot learning with retrieval augmented language models},
volume = {24},
issn = {1532-4435},
abstract = {Large language models have shown impressive few-shot results on a wide range of tasks. However, when knowledge is key for such results, as is the case for tasks such as question answering and fact checking, massive parameter counts to store knowledge seem to be needed. Retrieval-augmented models are known to excel at knowledge intensive tasks without the need for as many parameters, but it is unclear whether they work in few-shot settings. In this work we present Atlas, a carefully designed and pre-trained retrieval-augmented language model able to learn knowledge intensive tasks with very few training examples. We perform evaluations on a wide range of tasks, including MMLU, KILT and Natural Questions, and study the impact of the content of the document index, showing that it can easily be updated. Notably, Atlas reaches over 42\% accuracy on Natural Questions using only 64 examples, outperforming a 540B parameter model by 3\% despite having 50x fewer parameters.},
number = {1},
journal = {J. Mach. Learn. Res.},
author = {Izacard, Gautier and Lewis, Patrick and Lomeli, Maria and Hosseini, Lucas and Petroni, Fabio and Schick, Timo and Dwivedi-Yu, Jane and Joulin, Armand and Riedel, Sebastian and Grave, Edouard},
month = jan,
year = {2023},
keywords = {information retrieval, language models, retrieval augmented language models},
}
@misc{xu_search---chain_2024,
title = {Search-in-the-{Chain}: {Interactively} {Enhancing} {Large} {Language} {Models} with {Search} for {Knowledge}-intensive {Tasks}},
shorttitle = {Search-in-the-{Chain}},
url = {http://arxiv.org/abs/2304.14732},
doi = {10.48550/arXiv.2304.14732},
abstract = {Making the content generated by Large Language Model (LLM), accurate, credible and traceable is crucial, especially in complex knowledge-intensive tasks that require multi-step reasoning and each step needs knowledge to solve. Retrieval-augmented generation is good potential to solve this problem. However, where and how to introduce Information Retrieval (IR) to LLM is a big challenge. Previous work has the problems that wrong knowledge retrieved by IR misleads the LLM and interaction between IR and LLM breaks the reasoning chain of LLM. This paper proposes a novel framework named {\textbackslash}textbf\{Search-in-the-Chain\} (SearChain) for the interaction between LLM and IR to solve the challenges. First, LLM generates the reasoning chain named Chain-of-Query (CoQ) where each node consists of an IR-oriented query-answer pair. Second, IR verifies the answer of each node of CoQ. It corrects the answer that is not consistent with the retrieved information when IR gives high confidence, which improves the credibility. Third, LLM can indicate its missing knowledge in CoQ and rely on IR to provide this knowledge to LLM. These operations improve the accuracy in terms of reasoning and knowledge. Finally, SearChain generates the reasoning process and marks references to supporting documents for each reasoning step, which improves traceability. Interaction with IR in SearChain forms a novel reasoning path based on a tree, which enables LLM to dynamically modify the direction of reasoning. Experiments show that SearChain outperforms state-of-the-art baselines on complex knowledge-intensive tasks including multi-hop Q{\textbackslash}\&A, slot filling, fact checking, and long-form Q{\textbackslash}\&A.},
urldate = {2026-03-16},
publisher = {arXiv},
author = {Xu, Shicheng and Pang, Liang and Shen, Huawei and Cheng, Xueqi and Chua, Tat-Seng},
month = feb,
year = {2024},
note = {arXiv:2304.14732},
keywords = {Computer Science - Computation and Language},
}
@misc{zhu_static_2025,
title = {From {Static} to {Dynamic}: {A} {Streaming} {RAG} {Approach} to {Real}-time {Knowledge} {Base}},
shorttitle = {From {Static} to {Dynamic}},
url = {http://arxiv.org/abs/2508.05662},
doi = {10.48550/arXiv.2508.05662},
abstract = {Dynamic streams from news feeds, social media, sensor networks, and financial markets challenge static RAG frameworks. Full-scale indices incur high memory costs; periodic rebuilds introduce latency that undermines data freshness; naive sampling sacrifices semantic coverage. We present Streaming RAG, a unified pipeline that combines multi-vector cosine screening, mini-batch clustering, and a counter-based heavy-hitter filter to maintain a compact prototype set. We further prove an approximation bound {\textbackslash}\$E{\textbackslash}[R(K{\textbackslash}\_t)] {\textbackslash}ge R{\textasciicircum}{\textbackslash}* - L Δ{\textbackslash}\$ linking retrieval quality to clustering variance. An incremental index upsert mechanism refreshes prototypes without interrupting queries. Experiments on eight real-time streams show statistically significant gains in Recall{\textbackslash}@10 (up to 3 points, p {\textless} 0.01), end-to-end latency below 15 ms, and throughput above 900 documents per second under a 150 MB budget. Hyperparameter sensitivity analysis over cluster count, admission probability, relevance threshold, and counter capacity validates default settings. In open-domain question answering with GPT-3.5 Turbo, we record 3.2-point gain in Exact Match and 2.8-point gain in F1 on SQuAD; abstractive summarization yields ROUGE-L improvements. Streaming RAG establishes a new Pareto frontier for retrieval augmentation.},
urldate = {2026-03-16},
publisher = {arXiv},
author = {Zhu, Yuzhou},
month = jul,
year = {2025},
note = {arXiv:2508.05662},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Information Retrieval},
}
@inproceedings{prouteau_sinr_2021,
address = {Porto, Portugal},
series = {Lecture {Notes} in {Computer} {Science}},
title = {{SINr}: {Fast} {Computing} of {Sparse} {Interpretable} {Node} {Representations} is not a {Sin}!},
shorttitle = {{SINr}},
url = {https://hal.science/hal-03197434},
doi = {10.1007/978-3-030-74251-5_26},
abstract = {While graph embedding aims at learning low-dimensional representations of nodes encompassing the graph topology, word embedding focus on learning word vectors that encode semantic properties of the vocabulary. The first finds applications on tasks such as link prediction and node classification while the latter is systematically considered in natural language processing. Most of the time, graph and word embeddings are considered on their own as distinct tasks. However, word co-occurrence matrices, widely used to extract word embeddings, can be seen as graphs. Furthermore, most network embedding techniques rely either on a word embedding methodology (Word2vec) or on matrix factorization, also widely used for word embedding. These methods are usually computationally expensive, parameter dependant and the dimensions of the embedding space are not interpretable. To circumvent these issues, we introduce the Lower Dimension Bipartite Graphs Framework (LDBGF) which takes advantage of the fact that all graphs can be described as bipartite graphs, even in the case of textual data. This underlying bipartite structure may be explicit, like in coauthor networks. However, with LDBGF, we focus on uncovering latent bipartite structures, lying for instance in social or word co-occurrence networks, and especially such structures providing conciser and interpretable representations of the graph at hand. We further propose SINr, an efficient implementation of the LDBGF approach that extracts Sparse Interpretable Node Representations using community structure to approximate the underlying bipartite structure. In the case of graph embedding, our near-linear time method is the fastest of our benchmark, parameter-free and provides state-of-the-art results on the classical link prediction task. We also show that low-dimensional vectors can be derived from SINr using singular value decomposition. In the case of word embedding, our approach proves to be very efficient considering the classical similarity evaluation.},
urldate = {2026-03-16},
booktitle = {Lecture {Notes} in {Computer} {Science},},
publisher = {Springer, Cham},
author = {Prouteau, Thibault and Connes, Victor and Dugué, Nicolas and Perez, Anthony and Lamirel, Jean-Charles and Camelin, Nathalie and Meignier, Sylvain},
month = apr,
year = {2021},
keywords = {Community detection, Graph embedding, Linear-time algorithm, Link prediction, Network science, Word embedding},
pages = {325--337},
}
@article{hou_advancing_2025,
title = {Advancing continual lifelong learning in neural information retrieval: {Definition}, dataset, framework, and empirical evaluation},
volume = {687},
issn = {0020-0255},
shorttitle = {Advancing continual lifelong learning in neural information retrieval},
url = {https://www.sciencedirect.com/science/article/pii/S0020025524012829},
doi = {10.1016/j.ins.2024.121368},
abstract = {Continual learning refers to the capability of a machine learning model to learn and adapt to new information, without compromising its performance on previously learned tasks. Although several studies have investigated continual learning methods for neural information retrieval (NIR) tasks, a well-defined task definition is still lacking, and it is unclear how typical learning strategies perform in this context. To address this challenge, a systematic task definition of continual NIR is presented, along with a multiple-topic dataset that simulates continuous information retrieval. A comprehensive continual neural information retrieval framework consisting of typical retrieval models and continual learning strategies is then proposed. Empirical evaluations illustrate that the proposed framework can successfully prevent catastrophic forgetting in neural information retrieval and enhance performance on previously learned tasks. The results also indicate that embedding-based retrieval models experience a decline in their continual learning performance as the topic shift distance and dataset volume of new tasks increase. In contrast, pretraining-based models do not show any such correlation. Adopting suitable learning strategies can mitigate the effects of topic shift and data augmentation in continual neural information retrieval.},
urldate = {2026-03-16},
journal = {Information Sciences},
author = {Hou, Jingrui and Cosma, Georgina and Finke, Axel},
month = jan,
year = {2025},
keywords = {Catastrophic forgetting, Continual learning, Data augmentation, Neural information retrieval, Topic shift},
pages = {121368},
}
@misc{ni_towards_2025,
title = {Towards {Trustworthy} {Retrieval} {Augmented} {Generation} for {Large} {Language} {Models}: {A} {Survey}},
shorttitle = {Towards {Trustworthy} {Retrieval} {Augmented} {Generation} for {Large} {Language} {Models}},
url = {http://arxiv.org/abs/2502.06872},
doi = {10.48550/arXiv.2502.06872},
abstract = {Retrieval-Augmented Generation (RAG) is an advanced technique designed to address the challenges of Artificial Intelligence-Generated Content (AIGC). By integrating context retrieval into content generation, RAG provides reliable and up-to-date external knowledge, reduces hallucinations, and ensures relevant context across a wide range of tasks. However, despite RAG's success and potential, recent studies have shown that the RAG paradigm also introduces new risks, including robustness issues, privacy concerns, adversarial attacks, and accountability issues. Addressing these risks is critical for future applications of RAG systems, as they directly impact their trustworthiness. Although various methods have been developed to improve the trustworthiness of RAG methods, there is a lack of a unified perspective and framework for research in this topic. Thus, in this paper, we aim to address this gap by providing a comprehensive roadmap for developing trustworthy RAG systems. We place our discussion around five key perspectives: reliability, privacy, safety, fairness, explainability, and accountability. For each perspective, we present a general framework and taxonomy, offering a structured approach to understanding the current challenges, evaluating existing solutions, and identifying promising future research directions. To encourage broader adoption and innovation, we also highlight the downstream applications where trustworthy RAG systems have a significant impact.},
urldate = {2026-02-03},
publisher = {arXiv},
author = {Ni, Bo and Liu, Zheyuan and Wang, Leyao and Lei, Yongjia and Zhao, Yuying and Cheng, Xueqi and Zeng, Qingkai and Dong, Luna and Xia, Yinglong and Kenthapadi, Krishnaram and Rossi, Ryan and Dernoncourt, Franck and Tanjim, Md Mehrab and Ahmed, Nesreen and Liu, Xiaorui and Fan, Wenqi and Blasch, Erik and Wang, Yu and Jiang, Meng and Derr, Tyler},
month = feb,
year = {2025},
note = {arXiv:2502.06872},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language},
}
@misc{fan_survey_2024,
title = {A {Survey} on {RAG} {Meeting} {LLMs}: {Towards} {Retrieval}-{Augmented} {Large} {Language} {Models}},
shorttitle = {A {Survey} on {RAG} {Meeting} {LLMs}},
url = {http://arxiv.org/abs/2405.06211},
doi = {10.48550/arXiv.2405.06211},
abstract = {As one of the most advanced techniques in AI, Retrieval-Augmented Generation (RAG) can offer reliable and up-to-date external knowledge, providing huge convenience for numerous tasks. Particularly in the era of AI-Generated Content (AIGC), the powerful capacity of retrieval in providing additional knowledge enables RAG to assist existing generative AI in producing high-quality outputs. Recently, Large Language Models (LLMs) have demonstrated revolutionary abilities in language understanding and generation, while still facing inherent limitations, such as hallucinations and out-of-date internal knowledge. Given the powerful abilities of RAG in providing the latest and helpful auxiliary information, Retrieval-Augmented Large Language Models (RA-LLMs) have emerged to harness external and authoritative knowledge bases, rather than solely relying on the model's internal knowledge, to augment the generation quality of LLMs. In this survey, we comprehensively review existing research studies in RA-LLMs, covering three primary technical perspectives: architectures, training strategies, and applications. As the preliminary knowledge, we briefly introduce the foundations and recent advances of LLMs. Then, to illustrate the practical significance of RAG for LLMs, we systematically review mainstream relevant work by their architectures, training strategies, and application areas, detailing specifically the challenges of each and the corresponding capabilities of RA-LLMs. Finally, to deliver deeper insights, we discuss current limitations and several promising directions for future research. Updated information about this survey can be found at https://advanced-recommender-systems.github.io/RAG-Meets-LLMs/},
urldate = {2026-01-09},
publisher = {arXiv},
author = {Fan, Wenqi and Ding, Yujuan and Ning, Liangbo and Wang, Shijie and Li, Hengyun and Yin, Dawei and Chua, Tat-Seng and Li, Qing},
month = jun,
year = {2024},
note = {arXiv:2405.06211},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language, Computer Science - Information Retrieval},
}
@article{lewis_retrieval-augmented_2020,
title = {Retrieval-{Augmented} {Generation} for {Knowledge}-{Intensive} {NLP} {Tasks}},
journal = {arXiv: Computation and Language},
author = {Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and Küttler, Heinrich and Lewis, Mike and Yih, Wen-tau and Rocktäschel, Tim and Riedel, Sebastian and Kiela, Douwe},
year = {2020},
}
@article{liu_hm-rag_2025,
title = {{HM}-{RAG}: {Hierarchical} {Multi}-{Agent} {Multimodal} {Retrieval} {Augmented} {Generation}},
doi = {10.48550/arxiv.2504.12330},
journal = {arXiv.org},
author = {Liu, Pei and Liu, Xin and Yao, Ruoyu and Liu, Junming and Meng, Siyuan and Wang, Ding and Ma, Jun},
year = {2025},
}
@misc{touvron_llama_2023,
title = {{LLaMA}: {Open} and {Efficient} {Foundation} {Language} {Models}},
shorttitle = {{LLaMA}},
url = {http://arxiv.org/abs/2302.13971},
doi = {10.48550/arXiv.2302.13971},
abstract = {We introduce LLaMA, a collection of foundation language models ranging from 7B to 65B parameters. We train our models on trillions of tokens, and show that it is possible to train state-of-the-art models using publicly available datasets exclusively, without resorting to proprietary and inaccessible datasets. In particular, LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B. We release all our models to the research community.},
urldate = {2025-12-18},
publisher = {arXiv},
author = {Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and Martinet, Xavier and Lachaux, Marie-Anne and Lacroix, Timothée and Rozière, Baptiste and Goyal, Naman and Hambro, Eric and Azhar, Faisal and Rodriguez, Aurelien and Joulin, Armand and Grave, Edouard and Lample, Guillaume},
month = feb,
year = {2023},
note = {arXiv:2302.13971},
keywords = {Computer Science - Computation and Language},
}
@inproceedings{bell_its_2022,
address = {Seoul Republic of Korea},
title = {It’s {Just} {Not} {That} {Simple}: {An} {Empirical} {Study} of the {Accuracy}-{Explainability} {Trade}-off in {Machine} {Learning} for {Public} {Policy}},
isbn = {9781450393522},
shorttitle = {It’s {Just} {Not} {That} {Simple}},
url = {https://dl.acm.org/doi/10.1145/3531146.3533090},
doi = {10.1145/3531146.3533090},
language = {en},
urldate = {2025-09-11},
booktitle = {2022 {ACM} {Conference} on {Fairness} {Accountability} and {Transparency}},
publisher = {ACM},
author = {Bell, Andrew and Solano-Kamaiko, Ian and Nov, Oded and Stoyanovich, Julia},
month = jun,
year = {2022},
pages = {248--266},
}
@inproceedings{garouani_investigating_2024,
title = {Investigating the {Duality} of {Interpretability} and {Explainability} in {Machine} {Learning}},
url = {http://arxiv.org/abs/2503.21356},
doi = {10.1109/ICTAI62512.2024.00125},
abstract = {The rapid evolution of machine learning (ML) has led to the widespread adoption of complex "black box" models, such as deep neural networks and ensemble methods. These models exhibit exceptional predictive performance, making them invaluable for critical decision-making across diverse domains within society. However, their inherently opaque nature raises concerns about transparency and interpretability, making them untrustworthy decision support systems. To alleviate such a barrier to high-stakes adoption, research community focus has been on developing methods to explain black box models as a means to address the challenges they pose. Efforts are focused on explaining these models instead of developing ones that are inherently interpretable. Designing inherently interpretable models from the outset, however, can pave the path towards responsible and beneficial applications in the field of ML. In this position paper, we clarify the chasm between explaining black boxes and adopting inherently interpretable models. We emphasize the imperative need for model interpretability and, following the purpose of attaining better (i.e., more effective or efficient w.r.t. predictive performance) and trustworthy predictors, provide an experimental evaluation of latest hybrid learning methods that integrates symbolic knowledge into neural network predictors. We demonstrate how interpretable hybrid models could potentially supplant black box ones in different domains.},
urldate = {2025-08-21},
booktitle = {2024 {IEEE} 36th {International} {Conference} on {Tools} with {Artificial} {Intelligence} ({ICTAI})},
author = {Garouani, Moncef and Mothe, Josiane and Barhrhouj, Ayah and Aligon, Julien},
month = oct,
year = {2024},
note = {arXiv:2503.21356 [cs]},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Machine Learning},
pages = {861--867},
}
\ No newline at end of file
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment