Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability
Extracted source text · arXiv 2506.02138
\pdfoutput=1
\documentclass{article}
% if you need to pass options to natbib, use, e.g.:
\PassOptionsToPackage{numbers, compress}{natbib}
% before loading neurips_2025
% ready for submission
\usepackage[preprint]{neurips_2025}
\setlength{\textfloatsep}{12pt plus 2.0pt minus 2.0pt}
\newcommand{\myparagraph}[1]{\noindent{\textbf{#1}~~}}
% to compile a preprint version, e.g., for submission to arXiv, add add the
% [preprint] option:
% \usepackage[preprint]{neurips_2025}
% to compile a camera-ready version, add the [final] option, e.g.:
% \usepackage[final]{neurips_2025}
% to avoid loading the natbib package, add option nonatbib:
% \usepackage[nonatbib]{neurips_2025}
\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc} % use 8-bit T1 fonts
\usepackage{hyperref} % hyperlinks
\usepackage{url} % simple URL typesetting
\usepackage{booktabs} % professional-quality tables
\usepackage{amsfonts} % blackboard math symbols
\usepackage{nicefrac} % compact symbols for 1/2, etc.
\usepackage{microtype} % microtypography
\usepackage{xcolor} % colors
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{booktabs}
\usepackage{bm}
\usepackage{fancybox}
\usepackage{subcaption}
\usepackage{wrapfig}
\usepackage{amsthm}
\newtheorem{lemma}{Lemma}[section]
\newtheorem{example}{Example}[section]
\newcommand{\lw}[1]{\textcolor{black}{#1}}
\newcommand{\hc}[1]{\textcolor{black}{Hila: #1}}
\newcommand{\yb}[1]{\textcolor{black}{#1}}
\title{Revisiting LRP: Positional Attribution as the Missing
Ingredient for Transformer Explainability}
\author{%
Yarden Bakish \\
\small Tel-Aviv University\\
%\small \texttt{zimerman1@mail.tau.ac.il} \\
\And
Itamar Zimerman \\
\small Tel-Aviv University \\
%\small \texttt{zimerman1@mail.tau.ac.il} \\
\And
Hila Chefer \\
\small Tel-Aviv University \\
%\small \texttt{zimerman1@mail.tau.ac.il} \\
\And
Lior Wolf \\
\small Tel-Aviv University \\
%\small \texttt{zimerman1@mail.tau.ac.il}
}
% \author{%
% David S.~Hippocampus\thanks{Use footnote for providing further information
% about author (webpage, alternative address)---\emph{not} for acknowledging
% funding agencies.} \\
% Department of Computer Science\\
% Cranberry-Lemon University\\
% Pittsburgh, PA 15213 \\
% \texttt{hippo@cs.cranberry-lemon.edu} \\
% examples of more authors
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \AND
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
\begin{document}
\maketitle
\begin{abstract}
The development of effective explainability tools for Transformers is a crucial pursuit in deep learning research. One of the most promising approaches in this domain is Layer-wise Relevance Propagation (LRP), which propagates relevance scores backward through the network to the input space by redistributing activation values based on predefined rules. However, existing LRP-based methods for Transformer explainability entirely overlook a critical component of the Transformer architecture: its positional encoding (PE), resulting in violation of the conservation property, and the loss of an important and unique type of relevance, which is also associated with structural and positional features. To address this limitation, we reformulate the input space for Transformer explainability as a set of position-token pairs%, rather than relying solely on the standard vocabulary space
. This allows us to propose specialized theoretically-grounded LRP rules designed to propagate attributions across various positional encoding methods, including Rotary, Learnable, and Absolute PE. Extensive experiments with {both fine-tuned classifiers and zero-shot foundation models, such as LLaMA 3,} demonstrate that our method significantly outperforms the state-of-the-art in both vision and NLP explainability tasks. Our code is publicly available.
\vspace{0.5em}
\hspace{.5em}
\includegraphics[width=1.25em,height=1.15em]{figures/github.png}\hspace{.75em}
\parbox{\dimexpr\linewidth-7\fboxsep-7\fboxrule}{\url{https://github.com/YardenBakish/PE-AWARE-LRP}}
\vspace{-.5em}
\end{abstract}
\vspace{-7pt}
\section{Introduction\label{sec:intro}}
\vspace{-5pt}
Explainable AI (XAI) is increasingly vital in deep learning (DL), where models often achieve remarkable performance but operate as opaque ``black boxes''~\citep{arrieta2020explainable,das2020opportunities}. This lack of transparency reduces trust, limits user engagement, and complicates troubleshooting, thereby restricting the use of DL models in applications where decision-making transparency is essential. Consequently, developing XAI techniques for DL models has become an important research domain~\citep{samek2017explainable}. This task, however, is challenging, due to the inherent complexity of these models, which cannot be easily represented by simple functions.
Transformer-based architectures, which have become dominant in DL, present additional challenges for explainability due to their large scale, often containing billions of parameters. To address this, researchers have developed various attribution methods specifically designed for Transformers~\citep{chefer2021transformer,achtibat2024attnlrp,ali2022xai,abnar2020quantifying}. Among these, model-specific XAI techniques have gained prominence, providing explanations based on the model’s parameters, internal representations, and overall architecture.
The most effective model-specific XAI techniques, and the current state-of-the-art for Transformer explainability, are LRP-based, such as~\citep{achtibat2024attnlrp}. LRP is a well-established attribution technique that explains a model’s predictions by propagating relevance scores backward through the network, redistributing activation values based on predefined propagation rules. Unlike gradient-based methods, which often suffer from issues like vanishing gradients or numerical instabilities, LRP provides a more stable and precise way to trace how information flows through each layer.
\begin{wrapfigure}{r}{0.51\textwidth}
\centering
\centering
% \vspace{-7pt}
\begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
\includegraphics[width=0.12\textwidth]{figures/vision/img1/196_basic_custom_lrp_SEMANTIC_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img5/165_basic_custom_lrp_SEMANTIC_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img17/img17_sem.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img6/157_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\multicolumn{4}{c}{(a)} \\
\includegraphics[width=0.12\textwidth]{figures/vision/img1/196_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img5/165_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img17/img17_pos.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img6/157_basic_custom_lrp_PE_ONLY.png}
\\
\multicolumn{4}{c}{(b)} % \bottomrule
% \multicolumn{4}{c}{ \includegraphics[width=0.5\textwidth]{figures/conservation.png}}\\
% \multicolumn{4}{c}{(c)} \\
\end{tabular}
% \vspace{-4pt}
\caption{\textbf{(a)} Explainability heatmaps by the state-of-the-art Attention-LRP method (AttnLRP)~\citep{achtibat2024attnlrp}. \textbf{(b)} The LRP heatmap obtained directly from our PE-aware LRP rules. % our method for positional encoding (PE). %encoding (PE).
The relevancy captured by the PE is less fragmented and captures more of the object.}
\vspace{-13pt}
\label{fig:Motivation}
\end{wrapfigure}
%
Recently, several refinements have been proposed to improve the stability and faithfulness of LRP rules for Transformers, leading to more robust and reliable interpretability techniques. Notable examples include~\citep{chefer2021transformer,ali2022xai} and \citep{achtibat2024attnlrp}, which introduce custom rules for propagating LRP through attention mechanisms, layer normalization, and other key components. Despite these advancements, we identify a critical gap in this extensive line of work: all existing LRP-based methods for Transformers overlook the need for PE-aware LRP rules and do not propagate attribution through positional encoding. This omission results in the loss of a key aspect of relevancy related to positional and structural concepts, limiting the ability to provide faithful and comprehensive explanations.
%\begin{figure}[t]
To mitigate this problem, we propose Positional-Aware LRP (PA-LRP), a novel technique that significantly improves upon previous methods through two fundamental modifications: (i) Reformulating the input space of the Transformer explainability problem to incorporate positional information. Instead of relying solely on the vocabulary space in NLP and the patch space in vision, we define the input space as a set of position-token pairs. (ii) Introducing the first LRP rules specifically designed to propagate relevance across standard positional encoding (PE) layers, including learnable PE, Rotary PE~\citep{su2024roformer}, and others. To enhance stability and faithfulness, our rules are further improved through techniques such as reparameterization of PE layers, linearization, and defining an appropriate sink for positional relevance to ensure that position-associated information is properly absorbed, which we validate to be crucial for precise propagation.
Moreover, we provide complementary theoretical analysis, proving that our rules both satisfy the conservation property, and able to derive explanations more faithfully.
\textbf{Our main contributions}
consist of the following: (i) We identify a critical gap in current LRP-based XAI techniques for Transformers: they overlook the attribution of positional encodings (PE). This omission results in a violation of the conservation property for input-level PE, as shown in Lemma~\ref{lem:lemma1}, and leads to unfaithful heatmaps when handling positional features, as demonstrated in Lemma~\ref{lem:lemma3}. We empirically validate that this omission is a critical limitation by significantly outperforming existing methods, and demonstrating that in certain cases, assigning relevance to PE alone can surpass standard state-of-the-art Transformer explainability techniques, showing that this signal is significant%, as shown in Tables~\ref{tab:perturbationZeroShot}--~\ref{tab:pertubatoinNLp} and Figure~\ref{fig:qualitative_cmp_vision}
.
Additionally, the obtained signal is complementary and distinct from the non-positional signal, better capturing spatial, positional, and structural relationships, as shown in Figure~\ref{fig:Motivation}. (ii) We introduce PA-LRP, a theoretically grounded and PE-aware technique for assigning relevance in Transformers. As shown in Tables~\ref{tab:pertubatoinNLp}--~\ref{tab:perturbationZeroShot}, PA-LRP significantly outperforms previous methods across both fine-tuned classifiers and zero-shot foundation models, in both NLP and vision tasks. (iii) Providing an open-source and user-friendly implementation of our method, along with demos and practical examples, to facilitate adoption by the broader research and practitioner community.
\vspace{-4pt}
\section{Background and Related Work\label{sec:relatedWork}}
\vspace{-4pt}
In this section, we describe the scientific context for discussing LRP-based Transformer explainability, along with the necessary terminology and symbols needed to describe our method.%
%
\vspace{-3pt}
\subsection{Positional Encoding in Transformers}
\vspace{-3pt}
Transformer-based ~\citep{vaswani2017attention} architectures rely on self-attention, which computes contextual relationships between tokens using:
%
{\small
\begin{equation}
\text{Attention}(X) = \text{Softmax} \left( \frac{QK^T}{\sqrt{d_k}} \right) V,\quad Q = XW_Q,\quad K = XW_K,\quad V = XW_V
\end{equation}
}
%
%where, $K=XW_K, Q=XW_Q, V=XW_V$
where, $K, Q, V$ represent key, query, and value matrices respectively, $d_k$ is the embedding dimension, and $W_Q, W_K, W_V$ are learnable linear projection matrices. This attention mechanism is duplicated over several ``heads'' and is wrapped by standard DL peripherals such as Layer Normalization, FFNs, and skip connections, forming the core structure of a Transformer model by:
{\small
\begin{equation}%\nonumber
X' = \text{LayerNorm} \left( X + \text{Attention}(X) \right), \quad
% \end{equation}
% \begin{equation}\nonumber
X'' = \text{LayerNorm} \left( X' + \text{FFN}(X') \right) \,.
\end{equation}
}
where \textit{FFN} applies a two-layer linear transformation with activations in the middle.
%\subsection{Positional Encoding in Transformers}
Transformers operate on sets of tokens rather than ordered sequences, making them permutation-invariant by design. Unlike architectures with built-in order sensitivity such as RNNs~\citep{hochreiter1997long,jordan1997serial}, Transformers require explicit positional encoding (PE) to capture sequence structure. PE can be introduced at different stages of the model: it can be added to token embeddings at the input layer, as seen in learnable PE and sinusoidal PE~\citep{vaswani2017attention}, or integrated within the attention mechanism at each layer, as employed in Rotary PE (RoPE)~\citep{su2024roformer} and Alibi~\citep{presstrain}. The key insight of this paper is that while PE is well known for its important role in the forward pass~\citep{dufter2022position}, its crucial role in propagation-based XAI methods, such as LRP, has been largely overlooked, leading to violations of conservation and the loss of significant relevance, which often carries distinctive positional and structural meanings.
\noindent\noindent\textbf{Learnable PE.\quad} Learnable PE represents positions as trainable parameters, allowing the model to learn position representations directly from data. This approach offers flexibility and adaptability.
\noindent\noindent\textbf{Sinusoidal PE.\quad} Sinusoidal PE, as applied by the original Transformer model~\citep{vaswani2017attention}, encodes positions using sine and cosine functions with different non-trainable frequencies. Because it is based on absolute positions, it is less effective in tasks where relative positional information is more important.
\noindent\textbf{Rotary PE (RoPE).\quad} RoPE~\citep{su2024roformer} incorporates positional information by rotating token embeddings in a structured manner, enabling the model to naturally encode relative positions. Specifically, each key and query vector is transformed using a per-position block-diagonal rotation matrix. Unlike learnable or sinusoidal PEs, RoPE encodes relative positional relationships through the multiplication of rotation matrices. Due to its effectiveness, many popular foundation models, including SAM2~\citep{ravi2024sam},Pythia~\citep{biderman2023pythia}, LLaMA~\citep{touvron2023llama}, Qwen~\citep{bai2023qwen}, Gemma~\citep{team2024gemma}, and others are built on top or RoPE.% as their primary PE method.
Other PE techniques, such as ALiBi~\citep{presstrain} and relative PEs~\citep{shaw2018self,raffel2020exploring}, are described in Appendix~\ref{app:addPE}.
\vspace{-3pt}
\subsection{Model-Specific XAI and LRP}
\vspace{-3pt}
Methods for explaining neural models have been extensively studied in the context of DNNs~\citep{zhang2021survey,linardatos2020explainable}, particularly in NLP~\citep{arras-etal-2017-explaining,yuan2021explaining} and computer vision~\citep{selvaraju2017grad,lang2021explaining}, and across various architectures including transformers~\citep{chefer2021generic,kokalj2021bert}, RNNs~\citep{arras2019explaining,zhang2021survey}, CNNs~\citep{ibrahim2023explainable,zhang2018visual}, state-space models~\citep{jafari2024mambalrp,ali2024hidden}, and others. A widely adopted strategy for this task is the use of model-specific techniques, which exploit the internal architecture and parameters of neural models to generate explanations. One notable method in this category is LRP~\citep{bach2015pixel,montavon2019layer}, which propagates relevance scores, denoted by $\mathcal{R}(\cdot)$, backwards through the network by redistributing activation values. Propagation relies on predefined rules and interactions between tokens.
\noindent\textbf{LRP.\quad}%
LRP is an evolution of gradient-based methods, such as Input $\times$ Gradient~\citep{shrikumar2017learning,baehrens2010explain}, which often suffer from issues like numerical instabilities and gradient shattering~\citep{balduzzi2017shattered}. LRP enhances backpropagation rules by enforcing two key principles: (i) the conservation property, which ensures that the total relevance is preserved across layers. Namely, for a layer $M$, where $Y=M(X)$, the relevance of the output $\mathcal{R}(Y)$ is equal to the relevance of the input $\mathcal{R}(X)$. (ii) The prevention of numerical instabilities during propagation. To achieve these goals, LRP rules are often derived from the Deep Taylor Decomposition principle~\citep{montavon2017explaining}, redistributing relevance scores at each layer based on the first-order Taylor expansion of the layer’s function.
\vspace{-5pt}
\subsection{XAI for Transformers}
\vspace{-4pt}
The first model-specific XAI methods for Transformers were based on attention maps~\citep{caron2021emerging,clark2019does}, leveraging attention scores to quantify the contribution of each token to others across layers. Building on this approach,~\citet{abnar2020quantifying} introduced the attention rollout technique, which aggregates attention matrices across multiple layers to provide a more holistic explanation. However,~\citet{jain2019attention} later demonstrated that attention-based techniques can be misleading, as attention scores do not always correlate with gradient-based feature importance measures or actual model behavior. To address these limitations,~\citet{chefer2021transformer} developed a hybrid XAI method that combines LRP scores with attention maps, marking a breakthrough in the field by improving attribution fidelity.
Alternatively, an extensive body of work focuses on model-agnostic methods~\citep{mosca2022shap,kokalj2021bert,cheng2025unifying}, however, these approaches often exhibit lower performance compared to model-specific techniques.
Purely LRP-based XAI methods for Transformers were first introduced in~\citep{voita2021analyzing} and later refined by~\citet{ali2022xai}, who developed custom LRP rules tailored for LayerNorm and attention layers to preserve conservation properties and ensure numerical stability. More recently,~\citet{achtibat2024attnlrp} further improved this approach by designing more faithful propagation rules for self-attention, achieving state-of-the-art performance in Transformer explainability. To the best of our knowledge, this represents the most advanced technique in the field and serves as our primary baseline.
Interestingly, despite extensive research in this area, none of these approaches propagate relevance through PE layers. This omission leads to a loss of significant relevance associated with positional and structural features, ultimately resulting in less faithful and holistic attributions.
\vspace{-5pt}
\section{Method\label{sec:method}}
\vspace{-5pt}
\begin{wrapfigure}{r}{0.51\textwidth}
\centering
\vspace{-38pt}
\includegraphics[width=0.91\linewidth]{figures/PropogationOverPE-TheXAITransformer-Forpaper-NIPS.jpg}
\vspace{-7pt}
\caption{\small Visualization of our method for propagating PE-associated relevance. Purple arrows indicate the forward path, while blue arrows represent the LRP propagation rules. Dashed arrows denote custom PE-aware rules defined in our method.}
\label{fig:methodFig}
\vspace{-20pt}
\end{wrapfigure}
In this section, we describe our PE-aware LRP rules. We first revise the input space used in the Transformer explainability problem in Section~\ref{subsec:reform}. Then, building upon this formulation, we define our custom LRP rules in Section~\ref{subsec:PELRP-INPUT} and Section~\ref{subsec:PELRP-ATTN}. Finally, in Section~\ref{subsec:theory}, we prove that our PE-aware LRP rules are theoretically grounded.
\vspace{-3pt}
\subsection{Reformulating the Input Space\label{subsec:reform}}
\vspace{-3pt}
To comprehensively attribute positional information, we must define a \textbf{sink} that absorbs PE-associated relevance. To this end, we reformulate the explainability problem for Transformers. Given a sequence of embedded tokens of length $L$, denoted by $E_1, \ldots, E_L$%, \forall i E_i \in \mathbb{R}^{D}$), where %$|V|$ is the number of tokens in the vocabulary $V$, and
, where $D$ is the embedding size%$(\forall i : E_i \in \mathbb{R}^{D})$
, previous methods traditionally define the input space $\mathcal{S}$ as %
%
\begin{equation}
\mathcal{S} = \{E_i \mid i \in [L], E_i \in \mathbb{R}^D\}\,.
\end{equation}
%\begin{equation}
%$ \mathcal{S} = \{E_i \mid E_i \in \mathbb{R}^D\}$,
%\end{equation}
%
%where $E_i$ is the $i$-th row of the embedding matrix $E$. %
%
In contrast, we reformulate the input space as token-position pairs, with positional features defined separately for each layer, as follows:
\begin{equation}\label{eq:newInputSpace}
\mathcal{S} = \Big{\{}\Big{(}E_i, (P_{i,1},P_{i,2},\ldots,P_{i,K})\Big{)} \mid i \in [L],\quad E_i \in \mathbb{R}^D,\quad P_{i,1},\ldots,P_{i,K} \in \mathbb{R}^{D'} \Big{\}}\,,
\end{equation}
where %L$ is the sequence length,
$D'$ is the dimension of the positional embeddings, and $K$ is the number of layers. Thus, in our formulation, each token in the input space consists of two ingredients: one representing the semantic embedding $E_i$ for all token indices $i\in [L]$, and the other representing the per-layer positional embedding $P_{i,k}$ for all layers $k\in [K]$.
The use of separate per-layer sinks for positional relevance ensures that the omission of certain positional features in one layer does not obscure or override essential features captured in other layers, and that important positional attributions are not discarded.
%
%
%For simplicity, we tie the embedding dimensions of both the semantical and positional two vectors.
% To stably and effectively propagate relevance over the elements of the input space $\mathcal{S}$ and to define the elements of $S$ as a sink for position-associated relevance, the semantic modification of $\mathcal{S}$ in Eq.~\ref{eq:newInputSpace} should induce a corresponding modification in the computation of the forward pass of the Transformer. In particular, each positional operation, such as PE, must be represented as a linear function over position indices $j \in [L]$.
%Such linear form can be obtained by representing the embedding layer as an operation that builds on multiplication with one-hot vectors rather than selection from the PE LUT (Look-Up Table).
% However, transformers operate on continuous input embeddings rather than discrete token indices. Nevertheless, it is common practice to omit relevance propagation through the embedding matrix, as it often leads to poor empirical results and violations of the conservation property. %
% Thus, we further reformulate the input space as follows: Let $e_i \in \mathbb{B}^{1 \times |V|}$ and $f_j \in \mathbb{B}^{1 \times L} $ be one-hot binary vectors with a value of $1$ on the i and j coordinates, respectively. The input space is defined as
% \begin{equation}\label{eq:newInputSpaceIndicators}
% \mathcal{S}' = \{(e_i,f_j) \mid i \in [V], j \in [L]\}
% \end{equation}
%which can be equivalent to $\mathcal{S}$ as $e_i \cdot E = E_i$ where $\cdot$ denoted a matrix multiplication.
Building upon the formulation of Eq.~\ref{eq:newInputSpace}, the next two sections define LRP rules that enable stable propagation of relevance across PE components. In Section~\ref{subsec:PELRP-INPUT}, we present our rules for input-level PE, while Section~\ref{subsec:PELRP-ATTN} outlines our rules for attention-level PE.
\vspace{-4pt}
\subsection{LRP-rules for Input Level PE\label{subsec:PELRP-INPUT}}
\vspace{-3pt}
We begin with the simplest form of positional encoding—learnable PE—and then demonstrate that other input-level PEs can be reparameterized in a similar manner. Since PE is incorporated only at the input layer, we assume for brevity that $P_i$ is a vector rather than a matrix, namely $P_i = P_{i,1}$. We also tie the embedding dimensions of both the semantical and positional vectors ($D=D'$).
%\noindent\textbf{Learnable PE.\quad}
\myparagraph{Learnable PE.} %
This layer learns positional information during training through a positional embedding matrix $P'\in \mathbb{R}^{L' \times D}$ where $D$ represents the embedding dimension of positional information, and $L'$ denotes the maximum sequence length. For each sample, the positional and semantic embeddings are summed to obtain the final input representation. Formally, the combined embedding for the token at position $i$ is given by $P'_i + E_i$ where $P'_i$ is the i-th row of the positional embedding matrix $P'$.
% Via Eq.~\ref{eq:newInputSpaceIndicators}, we can represent learnable PE as a \textbf{linear operation} by $f_i P = E_i$, and the entire embedding layer as:
% \begin{equation}\label{eq:EmbedLayer}
% \text{Embedding}(e_i,f_j) = z_j = e_i E + f_j P, \quad z_j \in \mathbb{R}^{1 \times D}
% \end{equation}
% \begin{equation}\label{eq:EmbedLayer}
% \text{Embedding}(e_i,f_j) = z_j = e_i E + f_j P, \quad z_j \in \mathbb{R}^{1 \times D}
% \end{equation}
Thus, we can propagate relevance from the input of the first transformer block of the $i$-th token $z_i$ denoted by $\mathcal{R}(z_i)$, to the positional component $P'_i$ by using the standard LRP-$\epsilon$ rule for addition \citep{achtibat2024attnlrp}: %
%
% Thus, since $f_j$ is a one-hot vector, we can propagate relevancy from the input of the first transformer block $\mathcal{R}(z_i)$ to the positional index $\mathcal{R}(j)$ through the positional embedding matrix $P$ by the following rules:
% \begin{equation}
% \text{PA-LRP for input-level PE}:\quad \mathcal{R}(j) = \frac{1}{2}\mathcal{R}(z_j)\cdot (f_j P)
% \end{equation}
{\vspace{-5pt}
\begin{equation}\label{eq:LernablePELRP}
\text{PA-LRP for input-level PE}:
\mathcal{R}(P'_i) = P'_i\frac{\mathcal{R}(z_i)}{P'_{i}+E_{i} +\epsilon}\,.
\end{equation}
}
%
% As positive and negative relevance can cancel each other out, we refine this rule by the following normalized variant:
% \begin{equation}\label{eq:PA-LRP-input-normalzied}
% \mathcal{R}(j) = \frac{1}{2{D}}|\mathcal{R}(z_j)|\cdot |(f_j P)|
% \end{equation}
% where $|\cdot|$ applies the absolute value to each coordinate of the vector, and we divide by ${D}$ to ensure that the position-associated relevance is \textbf{normalized} and does not increase with the embedding size.
%\noindent\textbf{Sinusoidal PE.\quad}
\myparagraph{Sinusoidal PE.}%
This method encodes position information via a unique vector of sine and cosine values constructed by:
\begin{equation}\label{eq:sinPE}
% \text{Sinusoidal PE(j)}[d] = \begin{cases}
% \sin\Big{(} \frac{j}{10000^{\frac{2i}{D}}} \Big{)} \quad\text{if d is even} \\
% \cos\Big{(} \frac{j}{10000^{\frac{2i+1}{D}}} \Big{)} \quad\text{otherwise}
% \end{cases}
\text{Sinusoidal PE(i)}[2d] = \sin\left(\frac{i}{10000^{\frac{2d}{D}}}\right),\quad
% \end{equation}
% \begin{equation}
\text{Sinusoidal PE(i)}[2d+1] = \cos\left(\frac{i}{10000^{\frac{2d}{D}}}\right)\,.
\end{equation}
%
Thus, the values derived from Eq.~\ref{eq:sinPE} can be used to reparameterize the positional embedding matrix $P'$, replacing the learned vectors with their corresponding sine and cosine values. Such reparameterization eliminates the need to propagate gradients through non-linear functions such as sine and cosine, improving efficiency and stability.
\vspace{-4pt}
\subsection{LRP-rules for Attention-level PE\label{subsec:PELRP-ATTN}}
\vspace{-3pt}
For attention-level PE, we focus on describing the PA-LRP rules for RoPE~\citep{su2024roformer} as a representative example. For the derivation of the PA-LRP rules for ALiBi~\citep{presstrain}, we refer the reader to Appendix~\ref{app:AlibiRules}.
%\noindent\textbf{RoPE.\quad}
%\myparagraph{RoPE.}%
At each layer $k$, RoPE modifies the queries ($Q$) and keys ($K$) matrices before computing the attention scores. This modification is done by multiplying each key and query vector by a position-dependent rotation matrix $R_{i,k} \in \mathbb{R}^{D \times D}$ where $i \in [L]$. The rotation matrix is a block-diagonal matrix defined as follows: %
% {\small
% \begin{equation}
% R_j =
% \begin{bmatrix}
% \cos \theta_j^{(1)} & -\sin \theta_j^{(1)} & 0 & 0 & \dots & 0 & 0 \\
% \sin \theta_j^{(1)} & \cos \theta_j^{(1)} & 0 & 0 & \dots & 0 & 0 \\
% 0 & 0 & \cos \theta_j^{(2)} & -\sin \theta_j^{(2)} & \dots & 0 & 0 \\
% 0 & 0 & \sin \theta_j^{(2)} & \cos \theta_j^{(2)} & \dots & 0 & 0 \\
% \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
% 0 & 0 & 0 & 0 & \dots & \cos \theta_j^{(D/2)} & -\sin \theta_j^{(D/2)} \\
% 0 & 0 & 0 & 0 & \dots & \sin \theta_j^{(D/2)} & \cos \theta_j^{(D/2)}
% \end{bmatrix}
% \end{equation}
{\small
\begin{equation}\label{eq:rotaryMatrices}
\forall i\in [L], k \in [K]: R_{i,k} =
\begin{bmatrix}
\cos \theta_i^{(1)} & -\sin \theta_i^{(1)} & \dots & 0 & 0 \\
\sin \theta_i^{(1)} & \cos \theta_i^{(1)} & \dots & 0 & 0 \\
% 0 & 0 & \cos \theta_i^{(2)} & -\sin \theta_i^{(2)} & \dots & 0 & 0 \\
% 0 & 0 & \sin \theta_i^{(2)} & \cos \theta_i^{(2)} & \dots & 0 & 0 \\
\vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & \dots & \cos \theta_i^{(D/2)} & -\sin \theta_i^{(D/2)} \\
0 & 0 & \dots & \sin \theta_i^{(D/2)} & \cos \theta_i^{(D/2)}
\end{bmatrix}
\end{equation}
}%
where each rotation angle \( \theta_i^{(m)} \) is defined as %follows:
%
%\begin{equation}\label{eq:rotaryAngels}
$\theta_i^{(m)} = i \omega_m$, where $\omega_i = 10000^{\frac{D}{2(m-1)}}$.
%\end{equation}
Note that in RoPE, as in other attention-level positional encodings, the positional information is represented by a matrix $R_{i,k}$, not a vector. Accordingly, we assume:
\begin{equation}\label{eq:flatten}
% $ P_{i,k} = \text{Flatten}(R_{i,k})$, $D' = D^2$.
P_{i,k} = \text{Flattening}(R_{i,k}), \quad D' = D^2\,.
\end{equation}
Thus, we can propagate relevance from the matrix $\mathcal{R}(R_{i,k})$ to the vector $\mathcal{R}(P_{i,k})$ by flattening it:
\begin{equation}\label{eq:flattenRelevance}
\mathcal{R}(P_{i,k}) = \text{Flattening}(\mathcal{R}(R_{i,k}))\,.
\end{equation}
%
Now, a key remaining step is to define how relevance should be propagated to $R_{i,k}$. The RoPE computation is executed before computing the attention scores, transforming the per-position queries and keys as follows: %
%
%
\begin{equation}\nonumber %\label{eq:rotaryModiftyKQ}
\forall i \in [L]%,k \in [K]
: \tilde{\bm{Q}}_i = R_{i,k} \bm{Q}_i, \quad \tilde{\bm{K}}_i = R_{i,k} \bm{K}_i,\quad
% \end{equation}
% \begin{equation}
\text{RoPE Attention}(X) = \text{Softmax} \left( \frac{\tilde{\bm{Q}}\tilde{\bm{K}}^T}{\sqrt{d_k}} \right) V\,.
\end{equation}
%
Our formulation builds on top of AttnLRP~\citep{achtibat2024attnlrp}, which propagates relevance over the queries $\tilde{\bm{Q}}$ and keys $\tilde{\bm{K}}$, resulting in their corresponding relevance scores $\mathcal{R}(\tilde{\bm{Q}})$,$\mathcal{R}(\tilde{\bm{K}})$.
To propagate relevance from these matrices to the rotation matrices $R_{i,k}$, we apply the LRP rule for matrix multiplication employed by~\citet{achtibat2024attnlrp} separately to the key and query matrices, and then sum the resulting terms to produce a final attribution map per attention layer, as follows:
%
%
%
% parts in the input space (see Eq.~\ref{eq:newInputSpaceIndicators}) faithfully (according to Eqs.~\ref{eq:rotaryModiftyKQ},~\ref{eq:rotaryAngels},~\ref{eq:rotaryMatrices}), we first embed the flattened vectors obtained from the rotation matrices into an embedding matrix $G \in \mathbb{R}^{L \times (\frac{D}{2})^2}$, such that $R_j = \text{reshape}(f_j G)$. %
% Now, we can define the PA-LRP rules for RoPE:
%
%
\begin{equation}\label{eq:PA-LRP-matmul}
\forall i \in [L]%, \forall k \in [K]
: \mathcal{R}(R_{i,k}) = \frac{1}{2}\mathcal{R}(\tilde{\bm{Q_{i}}}) + \frac{1}{2}\mathcal{R}(\tilde{\bm{K_{i}}})\,.
\end{equation}
Finally, we aggregate the relevance scores across K layers and across feature dimension D, summing only the positive contributions, following a strategy similar to that proposed by~\citet{chefer2021transformer} and~\citet{xiong2024explainableartificialintelligencexai}:%~\citet{chefer2021transformer}: %
%
\begin{equation}\label{eq:aggrPositive}
\mathcal{R}_i = \sum_{d \in [D]} E_i [d]^+ + \sum_{k \in [K]}\sum_{d' \in [D']} P_{i,k}[d]^+ \,,
\end{equation}
where $\mathcal{R}_i$ is the final relevance for token $i$, and $(\cdot)^+$ denotes the ReLU function, which filters out negative values. %
\myparagraph{Overall Method.} %
% \noindent\textbf{Overall Method.\quad}%
Our PA-LRP rules allow us to assign relevance to the positional part of the input space. For the non-positional part, we use the same rules as defined in AttnLRP~\citep{achtibat2024attnlrp}.
It is worth noting that, although our rules in Eqs.\ref{eq:LernablePELRP},\ref{eq:flattenRelevance},\ref{eq:PA-LRP-matmul}, and \ref{eq:aggrPositive} are built on top of the AttnLRP framework, they are not limited to it. Our input-level PE rules can be decoupled and applied to any LRP method, while the attention-level PE rules can be integrated with alternative formulations, as long as they propagate relevance through the attention matrices and preserve the connection between PE and the computational graph.
As a result, similar to other LRP methods, our approach can produce explainability maps with computational efficiency comparable to a single backward pass. We further clarify that although our method introduces several modifications in the forward path and input space, it does not require any changes to the transformer itself. Instead, these modifications propose an equivalent forward path that allows us to better define the propagation rules.
\vspace{-4pt}
\subsection{Theoretical Analysis\label{subsec:theory}}
\vspace{-3pt}
To support our PA-LRP rules, we provide theoretical evidence demonstrating that they satisfy the key LRP criteria. First, the following two lemmas prove that our proposed LRP rules satisfy the conservation property.
\begin{lemma}\label{lem:lemma1}
\textit{For input-level PE transformers, the conservation property is violated when disregarding the positional embeddings' relevancy scores.}
\end{lemma}
\begin{lemma}\label{lem:lemma2}
\textit{For attention‐level PE transformers, our PE‐LRP rules satisfy the conservation property.}
\end{lemma}
%
Next, we present a lemma based on a key example illustrating that existing methods exhibit low faithfulness. In particular, we show that in simplified settings, applying standard LRP techniques without incorporating position-aware LRP rules leads to unfaithful explanations in tasks that heavily depend on positional features. %, such as predicting the number of tokens.%
%
\begin{lemma}\label{lem:lemma3}
\textit{For attention‐level PE transformers, current LRP attribution rules achieve low faithfulness, especially when considering positional features.}
\end{lemma} %
%
The proofs and examples are detailed in Appendix~\ref{app:Proofs}.
%
\vspace{-4pt}
\section{Experiments}
\vspace{-4pt}
To assess the effectiveness of our PA-LRP rules, we conduct a comprehensive set
of experiments in both the NLP and vision domains. First, in Section~\ref{sec:resultsNLP}, we perform perturbation tests with both zero-shot foundation models and finetuned classifiers, as well as performing an ablation study. Next, in Section~\ref{sec:resultsVision}, we conduct perturbation and segmentation tests in the Vision domain using DeiT~\citep{touvron2021training}.%, including perturbation and segmentation tests.
%Finally, in Sec.~\ref{sec:resultsAblations}, we conduct several ablation studies and empirical analyses to quantifying the significance of our method and justify design choices.
We begin by describing our baselines, ablation variant, and evaluation metrics:
\noindent\textbf{Baseline and Ablation Variant.}\quad %
Our primary baseline for comparison is AttnLRP~\citep{achtibat2024attnlrp}, as it represents the SoTA in general transformer XAI, and our method builds on top of it for non-positional components. The key distinction between our approach and this baseline (as well as other LRP-based methods) is our ability to attribute relevance to positional information. Our composite approach that balances both positional and non-positional relevance is denoted as PA-LRP, or 'ours'. Additionally, to isolate the effect of the positional encoding, we introduce an ablation variant denoted by 'PE Only', which directly measures the relevance assigned to positional components at the input space using our custom attribution rules.
%\noindent\textbf{Metrics.}\quad %
Although empirical evaluation of attribution methods is inherently challenging, we validate our PA-LRP method using perturbation and segmentation tests. Below, we describe these metrics:% in detail:
\noindent\textbf{Perturbation Tests.}\quad %
Perturbation tests are split into two metrics: positive and negative perturbations, which differ in the order in which pixels or tokens are masked. In positive perturbation, pixels or tokens are masked in descending order of relevance. An effective explanation method identifies the most influential regions, leading to a noticeable drop in the model’s score (measured in comparison to the predicted or target class) as these critical areas are gradually removed. In negative perturbation, masking begins with the least relevant elements and progresses toward the more important ones. A reliable explanation should keep the model’s prediction stable, demonstrating robustness even when unimportant components are masked.
Following~\citep{ali2022xai,zimerman2025explaining}, in both Vision and image domains, the final metric is quantified using the Area-Under-Curve (AUC), capturing model accuracy relative to the percentage of masked pixels or tokens, from 10\% to 90\%.
\noindent\textbf{Segmentation Tests.}\quad%
For attribution methods in vision, segmentation tests are a set of evaluations used to assess the quality of a model’s ability to distinguish foreground from background in an image. %
These tests compare the labeled segmentation image, which indicates whether each pixel belongs to the background or the foreground, with the explainability map, after it has been binarized using a thresholding technique. Then, several metrics are computed over both images: (i) Pixel Accuracy: The percentage of correctly classified pixels, measuring how well the predicted segmentation aligns with the ground truth. (ii) Mean Intersection-over-Union (mIoU): The ratio of the intersection to the union of the predicted and ground-truth segmentation maps, averaged across all images. (iii) Mean Average Precision (mAP): A metric that considers precision and recall trade-offs at different thresholds, providing a robust assessment of segmentation quality.
\vspace{-4pt}
\subsection{Results in NLP\label{sec:resultsNLP}}
\vspace{-3pt}
For experiments in the NLP% domain
, we first present results for perturbation tests, including an ablation study. For our tests, we adopt settings defined in~\citep{ali2022video,zimerman2025explaining} and we present qualitative results in Appendix.~\ref{sec:additonalNLPExamples}.
\myparagraph{Perturbation Tests for Finetuned Models.} %
%\noindent\textbf{Perturbation Tests for NLP.\quad}%
We conduct perturbation tests on three LLMs, finetuned on the IMDB classification dataset: LLaMa 2-7B~\citep{touvron2023llama}, LLaMa 2-7B Quantized, and Tiny-LLaMa~\citep{zhang2024tinyllama}.
The results presented in Table~\ref{tab:pertubatoinNLp} demonstrate that our method achieves better scores than the LRP baseline across all metrics %and models
. In particular, our approach improves the AU-MSE score in the generation scenario by 14.5\% for LLaMa 2-7B, 10.6\% for LLaMa 2-7B Quantized, and 51.41\% for Tiny-LLaMa.
\begin{table}[t]\centering
\small
\setlength{\tabcolsep}{3pt}
\caption{\small \textbf{Perturbation Tests in NLP.} Evaluation of LLaMa-2 7B, Quantized LLaMa-2 7B, and Tiny-LLaMa, all finetuned on IMDB, on pruning and generation perturbation tasks. AttnLRP~\citep{achtibat2024attnlrp} is the LRP baseline. The metrics used are AUAC (area under activation curve, higher is better) and AU-MSE (area under the MSE, lower is better).\label{tab:pertubatoinNLp}}
\resizebox{0.66\textwidth}{!}{
\begin{tabular}{c c c c c c }
\toprule
\multicolumn{1}{c}{Model} & \multicolumn{1}{c}{Method} & \multicolumn{2}{c}{Generation} & \multicolumn{2}{c}{Pruning} \\
& & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$} & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$}\\
\midrule
% \multicolumn{5}{c}{LLaMa 2-7b}\\
%
LLaMa-2 7B & AttnLRP& 0.779& 7.629& 0.777& 6.548\\
LLaMa-2 7B & PE Only& 0.771& 6.792& 0.771& 6.823\\
LLaMa-2 7B & Ours& \textbf{0.796}& \textbf{6.521}& \textbf{0.790}& \textbf{6.325}\\
% % LLaMa 2-7b & LRP+PE+Reform& 0.792& 6.771& 0.786&6.542\\
\midrule
% \multicolumn{5}{c}{LLaMa 2-7b Quantized}\\
LLaMa-2 7B Quantized & AttnLRP& 0.774& 11.348& 0.767& 10.067\\
LLaMa-2 7B Quantized & PE Only& 0.758& 10.730& 0.758& 10.774\\
LLaMa-2 7B Quantized & Ours& \textbf{0.785}& \textbf{10.137}& \textbf{0.778}& \textbf{9.685}\\
% LLaMa 2-7B Quantized & LRP+PE+Reform& 0.782& 10.251& 0.777&9.706\\
%
\midrule
% \multicolumn{5}{c}{Tiny-LLaMa}\\
Tiny-LLaMa-2 7B & AttnLRP& 0.803& 8.065& 0.792& 4.030\\
Tiny-LLaMa-2 7B & PE Only& 0.788& \textbf{3.918}& 0.788& \textbf{3.947}\\
Tiny-LLaMa-2 7B & Ours& \textbf{0.806}& {4.915}& \textbf{0.805}& 4.082\\
%Tiny-LLaMa & LRP+PE+Reform& \textbf{0.810}& 5.175& \textbf{0.809}&4.005\\
\bottomrule\\
\end{tabular}
}
\vspace{-11pt}
% \caption{\textbf{Perturbation Tests in NLP.} Experiments conducted with LLaMa-2 7B ('LMA'), Quantized LLaMa-2 7B ('LMA Q.') and Tiny-LLaMa (T.-LMA). "AttnLRP represents the baseline by~\citet{achtibat2024attnlrp}.. The metrics are: (1) AUAC - area under activation curve}
% The metrics are (1) AUAC - area under activation curve. Higher AUAC is
% better and indicates that: (a) pruning - removing less relevant nodes has little effect on model prediction (b) generation - adding the most relevant nodes strongly
% activates the correct model prediction (2) AU-MSE - area under the mean squared error $(y_0 - y_t)^2$. Lower AU-MSE is better: (a) pruning - removing less
% relevant nodes has little effect on the model prediction (b) generation - adding the most relevant nodes strongly
% reduces mean squared error.
\end{table}
\myparagraph{Perturbation Tests in Zero-Shot Settings.}
We use LLaMa 3-8B ~\citep{grattafiori2024llama} to evaluate explainability performance in zero-shot setting.
The results presented in Table~\ref{tab:perturbationZeroShot} showcase the superiority of our method across all metrics. \textbf{(i) Multiple-Choice Question Answering (MCQA):} our approach improves, on both generation and pruning scenarios, the AUAC score by approximately 3.2\%, and AU-MSE score by approximately by 7.7\%.
%We achieve drastic improvement for MCQA, due to each question in the dataset consisting of relatively small number of tokens, making every slight shift in relevancy scores critical to the model's prediction. In Fig.~\ref{fig:qualitative_cmp_NLP_ARC} in the Appendix, we provide an example for how such sensitivity plays a key role during perturbation tests. Despite this complexity, we assert the validity of our method with a second set of experiments, as detailed below.
\textbf{(ii) Next Token Prediction:} our approach improves the AUAC score by approximately 0.5\% on both generation and pruning scenarios, and AU-MSE score by approximately by 3\% on both scenarios. In contrast to MCQA, the Wikipedia dataset consists relatively long texts, making shifts in relevancy distributions less critical to the model's prediction.
\begin{wraptable}{r}{0.50\textwidth}
\vspace{-9pt}
%\begin{table}[t]\centering
\small
\setlength{\tabcolsep}{3pt}
\caption{\small \textbf{Ablation Study:} Analyzing the contribution of the multi-sink mechanism via perturbation tests in NLP. The evaluation was conducted on %LLaMa-2-7B using
the IMDB dataset.\label{tab:pertubatoinAblationNLp}}
\vspace{-3pt}
\resizebox{0.5\textwidth}{!}{
\begin{tabular}{c c c c c }
\toprule
\multicolumn{1}{c}{Method} & \multicolumn{2}{c}{Generation} & \multicolumn{2}{c}{Pruning} \\
& \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$} & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$}\\
\midrule
% \multicolumn{5}{c}{LLaMa 2-7b}\\
Ours& \textbf{0.796}& \textbf{6.521}& \textbf{0.790}& \textbf{6.325} \\
w/o Multi-Sink & 0.759 & 7.124 & 0.758 & 7.158%7
\\
\bottomrule\\
\end{tabular}
}
\vspace{-17pt}
\end{wraptable}
%\end{table}
\myparagraph{Ablation.}
To better understand the contribution of our PA-LRP rules, we conduct perturbation tests for the method that attributes solely position-associated relevance. The results are presented in the second, fourth, and sixth rows of Table~\ref{tab:pertubatoinNLp}, and second row of Table~\ref{tab:perturbationZeroShot}, and are denoted by 'PE-Only.' Surprisingly, this method produces results similar to the AttnLRP baseline, demonstrating the importance of PE-associated relevance, which carries a significant part of the signal. Notably, in Table ~\ref{tab:pertubatoinNLp}, this variant achieves the best score on the AU-MSE metric for Tiny-LLaMA, reducing the error by 50\% compared to AttnLRP~\citep{achtibat2024attnlrp}. Moreover, in Table~\ref{tab:pertubatoinAblationNLp}, we ablate the contribution of our multi-sink approach, which relies on drawing solely positive contributions across layers, as we aim to prevent the loss of positional relevance due to influence of negative contributions in final layers. We evaluate explainability performance for binary classification %task
of LLaMa-2-7B, using the same perturbation metrics, and report that the multi-sink approach improves the results by 7\%.
\begin{table}[t]\centering
\small
\setlength{\tabcolsep}{4pt}
\caption{\small \textbf{Perturbation Tests in NLP (Zero-Shot).} Evaluation of LLaMa-3 8B in zero-shot on generation and pruning perturbation tasks for both multiple-choice question answering and Next-Token Prediction (NTP) settings. Metrics reported are AUAC (area under activation curve, higher is better) and AU-MSE (area under MSE, lower is better). ``AttnLRP'' refers to the LRP baseline~\citep{achtibat2024attnlrp}. 'G' for generation and 'P' for pruning.}
\resizebox{\textwidth}{!}{
\begin{tabular}{c
c c c c
c c c c}
\toprule
\multicolumn{1}{c}{} &
\multicolumn{4}{c}{Multiple-Choice Question Answering} &
\multicolumn{4}{c}{Next Token Prediction} \\
\cmidrule(lr){2-5} \cmidrule(lr){6-9}
Method
& G. AUAC $\uparrow$ & G. AU-MSE $\downarrow$ & P. AUAC $\uparrow$ & P. AU-MSE $\downarrow$
& G. AUAC $\uparrow$ & G. AU-MSE $\downarrow$ & P. AUAC $\uparrow$ & P. AU-MSE $\downarrow$ \\
\midrule
AttnLRP
& 0.365 & 66.399 & 0.354 & 68.856
& 0.559 & 41.704 & 0.559 & 42.003 \\
PE Only
& 0.374 & \textbf{61.014} & 0.364 & \textbf{63.141}
&0.557 & 40.538 & 0.556 & 40.800 \\
Ours & \textbf{0.377} & 61.285 & \textbf{0.368} & 63.424
& \textbf{0.562} & \textbf{40.474} & \textbf{0.561} & \textbf{40.735} \\
\bottomrule\\
\end{tabular}
}
\label{tab:perturbationZeroShot}
\vspace{-12pt}
\end{table}
% ~\\
%{\color{blue}
%\subsection{Ablation\label{sec:ablation}}
%To clarify the contribution of our method, we conduct a series of ablation studies. First, in Tables~\ref{tab:visionPertubatoin}--\ref{tab:perturbationZeroShot}, we isolate the impact of our PA-LRP rules by comparing the XAI heatmaps derived from positional relevance alone to those based on semantic relevance, using perturbation and segmentation tests across both NLP and Computer Vision datasets.
%Moreover, in Tab.~\ref{tab:pertubatoinAblationNLp}, we ablate the contribution of our multi-sink approach, which is designed to prevent the loss of positional relevance. As can be seen, on perturbation testes with LLaMa-2-7B evaluated with the IMDB dataset, the multi-sink approach improves the results by 7\%.
% \section{Positional Lens}
% 1. PE-ATTR Over Layers
% 2. Experiments with Permutations
% 3. Vision Examples
% 4. The amount of PE-ATTR
% }
% \begin{figure}[t]
% \centering
% \includegraphics[width=0.47\textwidth]{figures/conservation.png}
% \caption{{\color{red}Yadern please relace} \textbf{Analysis of the conservation property.} The figure reveals that a key reason for the lack of conservation in previous work is the overlooked attribution of PEs.}
% \label{fig:conservation}
% \end{figure}
% \begin{figure}[t]
% \centering
% %\begin{tabular}{cccc}
% % \begin{tabular}{@{}c@{}c@{}c@{}c@{}}
% \centering
% \includegraphics[width=0.4550\textwidth]{figures/NLP/conservation_bar_plot.png}
% \caption{\textbf{Analysis of the conservation property.} We measure the sum of relevance at three different positions in the model: the first, intermediate, and penultimate layers. All normalized by the relevance in the last layer. We assess both the positional relevance (in orange) and the non-positional relevance (in blue). The figure reveals that a key reason for the lack of conservation in previous work is the overlooked attribution of PEs. Experiments are conducted with LLaMA-2 7B, LLaMA-2 7B Quantized, and Tiny-LLaMA, all fine-tuned on the IMDB dataset.}
% \label{fig:conservation}
% \end{figure}
\vspace{-3pt}
\subsection{Results for Vision Transformers\label{sec:resultsVision}}
%\begin{figure}[t]
\begin{wrapfigure}{r}{0.48\textwidth}
\vspace{-45pt}
\centering
%\begin{tabular}{cccc}
% \begin{tabular}{@{}c@{}c@{}c@{}c@{}}
\begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
%\toprule
%\midrule
\includegraphics[width=0.12\textwidth]{figures/vision/img1/img_196.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img1/196_basic_custom_lrp.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img1/196_basic_custom_lrp_PE_ONLY.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img1/196_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.12\textwidth]{figures/vision/img5/img_165.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img5/165_basic_custom_lrp.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img5/165_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img5/165_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.12\textwidth]{figures/vision/img17/img17.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img17/img17_ours.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img17/img17_pos.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img17/img17_sem.png} \\
\includegraphics[width=0.12\textwidth]{figures/vision/img11/img_5.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img11/5_basic_custom_lrp.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img11/5_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img11/5_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.12\textwidth]{figures/vision/img4/img_155.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img4/155_basic_custom_lrp.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img4/155_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img4/155_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.12\textwidth]{figures/vision/img6/img_157.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img6/157_basic_custom_lrp.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img6/157_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.12\textwidth]
{figures/vision/img6/157_basic_custom_lrp_SEMANTIC_ONLY.png} \\
\includegraphics[width=0.12\textwidth]{figures/vision/N3/4111.png} & \includegraphics[width=0.12\textwidth]{figures/vision/N3/4111_ours.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/N3/4111_POSITIONAL.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/N3/4111_SEMANTIC.png}\\
\includegraphics[width=0.12\textwidth]{figures/vision/img13/img13.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img13/img13_ours.png} &
\includegraphics[width=0.12\textwidth]{figures/vision/img13/img13_pos.png} & \includegraphics[width=0.12\textwidth]{figures/vision/img13/img13_sem.png} \\(a) & (b) & (c) & (d)
%\bottomrule
\end{tabular}
\vspace{-6pt}
\caption{\small Results of different explanation methods for DeiT. (a) Input image. (b) PA-LRP (ours), which includes PE attribution. (c) PE only LRP, (d) AttnLRP~\citep{achtibat2024attnlrp}, which does not attribute relevancy to PE.}\label{fig:qualitative_cmp_vision}
\vspace{-44pt}
\end{wrapfigure}
For vision models, we present both quantitative and qualitative analysis. % , conducted on DeiT variants across three model sizes (tiny, small, and base).
%\subsubsection{Qualitative Analysis}%
\myparagraph{Qualitative Analysis.}
For qualitative analysis, we visualize the explainability maps obtained from our method, the AttnLRP~\citep{achtibat2024attnlrp} baseline, and the ablation variant that focuses exclusively on PE-associated relevance, denoted by PE Only. Additional examples with larger images are presented in appendix~\ref{sec:additonalExamples}.
Figure~\ref{fig:qualitative_cmp_vision} presents a comparative visualization of these maps. The results reveal three notable trends. \noindent\textbf{(i) Effectiveness of PE-associated relevance:} The maps from the ablation variant perform at the same level as the AttnLRP baseline. This finding highlights the strength of our method in identifying important signals that previous works have overlooked, underscoring the importance of our PA-LRP rules. \noindent\textbf{(ii) The uniqueness of PE-associated relevance:} The attributed signal derived solely from positional-associated relevance %via our PA-LRP rules (denoted by 'PE Only')
captures unique relationships, exhibiting clearer spatial and structural patterns. In particular, relevance is distributed across the entire object, especially in the snake, bird, and shark examples. In contrast, the baseline method, which does not propagate relevance through PEs, produces a sparser pattern that does not focus on the entire object but instead is highly selective to specific patches. One possible explanation is that positional-associated relevance better captures concepts related to position, structure, order, and broader regions within the image. \noindent\textbf{(iii) The importance of balancing:} It is evident that the maps obtained from the PE-associated method and the baseline are complementary, and their combination, extracted via our approach, provides the most robust explanations.
\myparagraph{Quantitative Analysis.}
%\subsubsection{Quantitative Analysis}%
Here, we present our quantitative results through perturbation and segmentation tests.
\noindent\textbf{Perturbation Tests in Vision.}\quad %
The results for perturbation tests are shown in Table~\ref{tab:visionPertubatoin}, where we compare our method against the attention LRP baseline and an ablation variant that focuses solely on maps obtained for positionally associated relevance (PE only). Experiments are conducted using three model sizes: Tiny, Small, and Base.
\begin{table}[h]
\centering
\setlength{\tabcolsep}{3pt}
\begin{minipage}[t]{0.435\textwidth}
\centering
\caption{\small Perturbation Tests for DeiT Variants on ImageNet. AUC results for predicted class. Higher (lower) is better for negative (positive).}
\small
\begin{tabular}{c c c c c c }
\toprule
\multicolumn{1}{c}{M. Size}& \multicolumn{1}{c}{Method}& \multicolumn{2}{c}{Negative $\uparrow$ } & \multicolumn{2}{c}{Positive $\downarrow$} \\
& & Predicted & Target & Predicted & Target \\
\midrule
Base & AttnLRP & 52.185 & 47.516 & 10.784 & \textbf{8.032} \\
Base & Ours & \textbf{54.970} & \textbf{50.174} & \textbf{9.918} & 9.237 \\
\midrule
Small & AttnLRP & 50.662 & 45.105 & 10.511 & 9.761 \\
Small & Ours & \textbf{53.482} & \textbf{47.948} & \textbf{9.135} & \textbf{8.477} \\
\midrule
Tiny & AttnLRP & 43.832 & 37.499 & \textbf{2.796} & \textbf{2.503} \\
Tiny & Ours & \textbf{50.126} & \textbf{42.567} & 3.579 & 3.214 \\
\bottomrule \\
\end{tabular}
\label{tab:visionPertubatoin}
\end{minipage}
\hfill
\begin{minipage}[t]{0.46\textwidth}
\centering
\small
\caption{\small Segmentation performance of DeiT variants on ImageNet segmentation~\citep{guillaumin2014imagenet}. 'M.' for model. Higher is better. }
\begin{tabular}{c c c c }
\toprule
M. Size & Method & Pixel Accuracy $\uparrow$ & mIoU $\uparrow$\\
\midrule
Base & AttnLRP & 72.204 & 50.100 \\
Base & Ours & \textbf{72.698} & \textbf{51.400} \\
\midrule
Small & AttnLRP & 72.114 & 50.000 \\
Small & Ours & \textbf{73.060} & \textbf{51.700} \\
\midrule
Tiny & AttnLRP & 74.815 & 52.850 \\
Tiny & Ours & \textbf{76.613} & \textbf{55.920} \\
\bottomrule \\
\end{tabular}
\label{tab:visionSegmentation}
\end{minipage}
\vspace{-5pt}
\end{table}
% \begin{table}[t]\centering
% % \small
% \setlength{\tabcolsep}{3pt}
% \begin{tabular}{c c c c c c }
% \toprule
% \multicolumn{1}{c}{M. Size}& \multicolumn{1}{c}{Method}& \multicolumn{2}{c}{Negative $\uparrow$ } & \multicolumn{2}{c}{Positive $\downarrow$} \\
% & & \multicolumn{1}{c}{Predicted} & \multicolumn{1}{c}{Target} & \multicolumn{1}{c}{Predicted} & \multicolumn{1}{c}{Target} \\
% \midrule
% % \multicolumn{5}{c}{DeiT-Base}\\
% Base & AttnLRP & 52.185 & 47.516 & 10.784 & \textbf{8.032} \\
% % Base & +PE & 54.239 & 49.435 & \textbf{9.310} & 8.713 \\
% % Base & PE Only & \textbf{56.686} & \textbf{51.286} & 11.843 & 11.151 \\
% Base &Ours & \textbf{54.970} & \textbf{50.174} & \textbf{9.918} & 9.237 \\
% \midrule
% Small & AttnLRP & 50.662 & 45.105 & 10.511 & 9.761 \\
% % Small & +PE & 52.485 & 47.032 & \textbf{9.001} & \textbf{8.349} \\
% % Small & PE Only & \textbf{54.799} & \textbf{48.809} & 9.810 & 9.077 \\
% Small & Ours & \textbf{53.482} & \textbf{47.948} & \textbf{9.135} & \textbf{8.477} \\
% \midrule
% Tiny & AttnLRP & 43.832 & 37.499 & \textbf{2.796} & \textbf{2.503} \\
% % Tiny & +PE & 48.685 & 41.450 & 3.198 & 2.859 \\
% % Tiny & PE Only & 48.382 & 40.605 & 6.548 & 5.824 \\
% Tiny&Ours & \textbf{50.126} & \textbf{42.567} & 3.579 & 3.214 \\
% \bottomrule
% \end{tabular}
% \caption{Perturbation Tests for DeiT Variants on ImageNet. We present the AUC results (percentages) for the predicted class. For positive perturbation lower is better, and for negative perturbation higher is better. 'M. Size' for models size, and "AttnLRP represents the baseline by~\citet{achtibat2024attnlrp}.}
% \label{tab:visionPertubatoin}
% \end{table}
Notably, our method outperforms the baseline by a significant margin. For instance, in negative perturbation of the predicted class, our method improves the performance by an average of 3.97 points across the three model sizes. However, in positive perturbation, our method lags behind the baseline in half of the cases, though by a small margin of at most 1.2 points.
\vspace{-5pt}
\section{Discussion: The Role of Attributing PEs}
\vspace{-5pt}
Our {theoretical and} empirical analysis suggests that both semantic and positional relevance are complementary, and combining them is essential to provide precise explanations. LRP-type attribution creates pixel-level heatmaps, but can we characterize and identify which {\em concepts} are attributed mainly by positional relevance versus semantic relevance?
We may expect, for example, that objects that are usually placed in specific contexts (boats on water, airplanes in the sky) would display a more significant PE component. %on the top part of images
%We hypothesize that the answer is yes. For example, in computer vision, when performing classification on ImageNet, sky-related classes such as airplanes, or clouds tend to appear in the upper part of images, potentially making their features position-dependent.
%Beyond this simple example, which relies on absolute positions, a more crucial aspect lies in relative PEs.
Much of this position context is relative. RoPE, for example, captures relative position through the matrix multiplication of two position-dependent rotation matrices, which plays a fundamental role in capturing spatial features in vision tasks (e.g., objects spanning across multiple patches) and when modeling relationships between words in the same sentence in NLP. In such cases, our PA-LRP rules can effectively attribute positional features, that are largely ignored by standard LRP methods. %As an example, Fig.~\ref{fig:qualitative_cmp_vision} shows that maps of positional relevancy better capture structural and spatial relationships.
%{\color{red}Finally, another important question arises: How should we balance positional and semantic relevance? Our method adopts the Half-Rule, assigning equal weight to both. However, it remains an open question whether this approach is optimal and whether certain tasks or regimes might benefit from prioritizing one type of relevance over the other. We leave this question for future work.}
\vspace{-6pt}
\section{Conclusions}\label{sec:conclusion}
\vspace{-6pt}
This paper explores the importance of assigning LRP scores to positional information, a crucial component of Transformers and LLMs. Our theoretical and empirical analysis demonstrates that positional-associated relevance carries a unique type of significance and can drastically improve XAI methods for attention models. %
Regarding limitations, we emphasize that our work focuses on designing new custom LRP rules to propagate relevance through PEs, leveraging the insight that this aspect has been previously overlooked. However, we do not extend this insight to redesign or systematically revisit existing LRP rules. Such a redesign could offer an opportunity to empirically and theoretically establish improved LRP rules for attention mechanisms and Transformer models.
\section{Acknowledgments}
This work was supported by a grant from the Tel Aviv University Center for AI and Data Science (TAD). This research was also supported by the Ministry of Innovation, Science \& Technology ,Israel (1001576154) and the Michael J. Fox
Foundation (MJFF-022407).
\newpage
\bibliographystyle{plainnat}
\bibliography{references}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\newpage
\appendix
\section{Background for Additional PEs\label{app:addPE}}
In this appendix, we introduce additional PEs beyond those presented in Section~\ref{sec:relatedWork}.
\noindent\textbf{Relative Positional Bias (RPB).\quad}%
Similar to Alibi, RPB~\citep{raffel2020exploring} modifies the attention scores by introducing a learnable bias term that depends on the relative distance between query and key tokens. For a query at position $i$ and a key at position $j$, the attention scores are adjusted as follows:
\begin{equation}
A_{i,j} = A_{i,j} + B(|i-j|)
\end{equation}
%
where $B(i-j)$ is a learned bias function that depends only on the relative position difference $(i-j)$, rather than the absolute positions.
\noindent\textbf{Attention with Linear Biases (ALiBi).\quad}%
ALiBi~\citep{presstrain} is a positional encoding method designed to help transformers generalize to longer sequences when trained on shorter ones. Instead of using explicit positional embeddings, ALiBi modifies attention scores directly by introducing a learned linear bias that penalizes attention weights based on token distance.
Specifically, for a query token at position $j$, Alibi adjusts the attention scores as follows:
\begin{equation}
A'_{i,j} = A_{i,j} + m (|i-j|)
\end{equation}
%
where $m$ is a learned or predefined slope that controls how quickly attention strength decays with distance. Different attention heads can use different slopes, enabling some heads to focus more on local interactions while others capture long-range dependencies.
\section{PA-LRP Rules for Alibi\label{app:AlibiRules}}
Recall the main modification in the ALiBi computation:
\begin{equation}
A'_{i,j} = A_{i,j} + P_{i,j}, \text{ where } P_{i,j}= m (|i-j|)
\end{equation}
Adopting the same approach presented for RoPE, given the relevancy scores of $A'_{i,j}$, denoted by $\mathcal{R}(A'_{i,j})$, we define specialized rules to propagate relevancy from $A'_{i,j}$ to the positional terms of ALiBi at each layer, namely, indices $i$ and $j$. We begin by distributing the relevancy scores between $A_{i,j}$ and $P_{i,j}$, using the standard LRP-$\epsilon$ rule for addition, giving us:
{\vspace{-5pt}
\begin{equation}
\mathcal{R}(P_{i,j}) = P_{i,j}\frac{\mathcal{R}(A'_{i,j})}{A_{i,j}+P_{i,j} +\epsilon}
\end{equation}
}
We proceed to propagate the relevancy scores $\mathcal{R}(P_{i,j})$ to the positional encoding $i$ and $j$ in a similar fashion to our rules for RoPE. We make the following observations: (i) $m$ is a constant, resulting in 100\% of the relevancy to propagate from $P_{i,j}$ to $|i-j|$. (ii) Since we are using auto-regressive models, we get that $i>j$, allowing us to ignore the absolute value function (iii) The standard LRP-$\epsilon$ rule for addition applies the same of subtraction, as we can express $i-j$ as $i+(-j)$, and also $-j =(-1)\cdot j$, and since $-1$ is constant, we propagate the entire relevancy to $j$. That gives us:
{\vspace{-5pt}
\begin{equation}
\mathcal{R}(i) = i\frac{\mathcal{R}(P_{i,j})}{i+(-j)+ \epsilon}\text{,} \quad \mathcal{R}(j) = \mathcal{R}(-j) = j\frac{\mathcal{R}(P_{i,j})}{i+(-j)+ \epsilon}
\end{equation}
}
From hereon we adhere to our PA-LRP rules, aggregating the relevance scores of the positional terms across all layers as employed in Section~\ref{subsec:PELRP-ATTN}.
\section{Additional Qualitative Results in Vision\label{sec:additonalExamples}}
In addition to Figure~\ref{fig:qualitative_cmp_vision}, we provide more examples in Figures \ref{fig:attVisExamples1} -~\ref{fig:attVisExamples3}. As previously explained, PE-associated relevance better highlights the entire object, and overcomes the issue of over-consideration of the foreground, where extremely high relevancy scores are produced for patches which are more concerned with semantics or common patterns, like a bird's beak in the first row in Figure~\ref{fig:attVisExamples2}.
\begin{figure*}[t]
\centering
%\begin{tabular}{cccc}
% \begin{tabular}{@{}c@{}c@{}c@{}c@{}}
\begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
%Input & Ours & PE Only & AttnLRP~\citep{achtibat2024attnlrp} \\
% \cmidrule(lr){1-4}
%\midrule
\includegraphics[width=0.20\textwidth]{figures/vision/img1/img_196.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img1/196_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img1/196_basic_custom_lrp_PE_ONLY.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img1/196_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img2/img_17.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img2/17_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img2/17_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img2/17_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img3/img_66.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img3/66_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img3/66_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img3/66_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img4/img_155.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img4/155_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img4/155_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img4/155_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img5/img_165.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img5/165_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img5/165_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img5/165_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img6/img_157.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img6/157_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img6/157_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]
{figures/vision/img6/157_basic_custom_lrp_SEMANTIC_ONLY.png}
\end{tabular}
\caption{\textbf{Additional Qualitative Results In Vision.} Results of different explanation methods for DeiT. (a) The input image. (b) PA-LRP (ours), which include PE relevancy attribution. (c) PE only LRP, (d) AttnLRP~\citep{achtibat2024attnlrp}, which does not attribute relevancy to PE.\label{fig:attVisExamples1}}
\end{figure*}
\begin{figure*}
\centering
%\begin{tabular}{cccc}
% \begin{tabular}{@{}c@{}c@{}c@{}c@{}}
\begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
% Input & Ours & PE Only & AttnLRP~\citep{achtibat2024attnlrp} \\
% \cmidrule(lr){1-4}
%\midrule
\includegraphics[width=0.2\textwidth]{figures/vision/img17/img17.png} &
\includegraphics[width=0.2\textwidth]{figures/vision/img17/img17_ours.png} &
\includegraphics[width=0.2\textwidth]{figures/vision/img17/img17_pos.png} & \includegraphics[width=0.2\textwidth]{figures/vision/img17/img17_sem.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img8/img_89.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img8/89_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img8/89_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img8/89_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img9/img_2.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img9/2_basic_custom_lrp.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img9/2_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img9/2_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img13/img13.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img13/img13_ours.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img13/img13_pos.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img13/img13_sem.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img11/img_5.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img11/5_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img11/5_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img11/5_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/N6/71.png} & \includegraphics[width=0.20\textwidth]{figures/vision/N6/71_ours.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/N6/71_POSITIONAL.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/N6/71_SEMANTIC.png}
\end{tabular}
\caption{\textbf{Additional Qualitative Results In Vision.} Results of different explanation methods for DeiT. (a) The input image. (b) PA-LRP (ours), which include PE relevancy attribution. (c) PE only LRP, (d) AttnLRP~\citep{achtibat2024attnlrp}, which does not attribute relevancy to PE.\label{fig:attVisExamples2}}
\end{figure*}
\begin{figure*}
\centering
%\begin{tabular}{cccc}
% \begin{tabular}{@{}c@{}c@{}c@{}c@{}}
\begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
% Input & Ours & PE Only & AttnLRP~\citep{achtibat2024attnlrp} \\
% \cmidrule(lr){1-4}
%\midrule
\includegraphics[width=0.20\textwidth]{figures/vision/img12/img12.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img12/img12_ours.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img12/img12_pos.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img12/img12_sem.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img10/img_136.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img10/136_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img10/136_basic_custom_lrp_PE_ONLY.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img10/136_basic_custom_lrp_SEMANTIC_ONLY.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img14/img14.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img14/img14_ours.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img14/img14_pos.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img14/img14_sem.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img15/img15.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img15/img15_ours.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img15/img15_pos.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img15/img15_sem.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img16/img16.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img16/img16_ours.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img16/img16_pos.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img16/img16_sem.png}
\\
\includegraphics[width=0.20\textwidth]{figures/vision/img7/img_92.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img7/92_basic_custom_lrp.png} &
\includegraphics[width=0.20\textwidth]{figures/vision/img7/92_basic_custom_lrp_PE_ONLY.png} & \includegraphics[width=0.20\textwidth]{figures/vision/img7/92_basic_custom_lrp_SEMANTIC_ONLY.png}
\end{tabular}
\caption{\textbf{Additional Qualitative Results In Vision.} Results of different explanation methods for DeiT. (a) The input image. (b) PA-LRP (ours), which include PE relevancy attribution. (c) PE only LRP, (d) AttnLRP~\citep{achtibat2024attnlrp}, which does not attribute relevancy to PE.\label{fig:attVisExamples3}}
\end{figure*}
\clearpage
\section{Additional Qualitative Results in NLP\label{sec:additonalNLPExamples}}
We present qualitative results for NLP in Figure~\ref{fig:qualitative_cmp_NLP}. It can be seen that our method demonstrates better results in highlighting tokens crucial for prediction, along with their surrounding context, emphasizing it's superiority to draw relevancy based on both semantics and positionally. In (b) we see that the amount of artifacts is reduced drastically, with more relevancy channeled to the tokens essential for prediction ("They should have been giving a tribute to Branagh for bringing us one of the greatest films of all time").
% \section{Additional NLP Examples}
\input{NLPSamples.tex}
\newpage
\clearpage
\section{Proofs of Lemmas\label{app:Proofs}}
{
\renewcommand{\thesection}{\arabic{section}}
\setcounter{lemma}{0}%result in 3
\setcounter{section}{3}
\begin{lemma}\label{lem:lemma1Appendix}
\textit{For input-level PE transformers, the conservation property is violated when disregarding the positional embeddings' relevancy scores.}
\end{lemma}
}
\begin{proof}[Proof of Lemma~\ref{lem:lemma1Appendix}]
Let $Z$ be our input representation to the first transformer layer, such that $Z = P + E$, where $P$ and $E$ are the token and positional embeddings, respectively. Let $L$ be the number of layers in our transformer. Following the conservation property, the sum of the relevancy scores at any given layer $l$ should uphold: \begin{equation}\sum \mathcal{R}^{(L)} =\sum \mathcal{R}^{(l)} =\sum \mathcal{R}^{(0)} = \sum R_{Z} = \sum (\mathcal{R}_{E} + \mathcal{R}_{P})\end{equation}
When ignoring $\mathcal{R}_{P}$, we get the final relevancy attribution map $\mathcal{R}_{input}$, such that: \begin{equation} \sum \mathcal{R}^{(l)} = \sum (\mathcal{R}_{E} + \mathcal{R}_{P}) \neq \mathcal{R}_{E} = \mathcal{R}_{input} \end{equation}
directly violating the conservation property rule
\end{proof}
{
\renewcommand{\thesection}{\arabic{section}}
\setcounter{lemma}{1}%result in 3
\setcounter{section}{3}
\begin{lemma}\label{lem:lemma2Appendix}
\textit{For attention‐level PE transformers, our PE‐LRP rules satisfy the conservation property.}
\end{lemma}
}
\begin{proof}[Proof of Lemma~\ref{lem:lemma2Appendix}]
Let $M$ be the number of layers in our Transformer, and $L$ the sequence length. We denote $\mathcal{R}^{(l)}$ as the relevancy score of the output at layer $l$. Beginning with $\mathcal{R}^{(M)}$ as the the model's output propagating relevancy backwards to achieve the final explanation map for the input embeddings $R_{E}$, we assume that the standard LRP method does not violate conservation, i.e:
\begin{equation} \forall l \in [M]: \quad \mathcal{R}^{(M)} = \mathcal{R}^{(l)} = \mathcal{R}_{E} \end{equation}
Recall that for our PA-LRP formulation, we achieve the final explanation map by summing together the semantic attribution $\mathcal{R}_{E}$, achieved by the standard LRP rules, and the positional relevancy $\mathcal{R}_{P}^{(l)}$ distributed across the absorbing sinks at each attention layer $l \in [M]$, giving us the final relevancy map $R_{E} + \sum_{l}\mathcal{R}_{P}^{(l)}$. We aim to prove the following:
\begin{equation}\label{eq:termToProveLemma3.2} \mathcal{R}^{(M)} =\mathcal{R}_{E} + \sum_{l}\mathcal{R}_{P}^{(l)} \end{equation}
Each attention layer in the transformer is computed using rotary attention:
{\vspace{-2pt}
\begin{equation}
\forall i \in [L], k \in [K]: \tilde{\bm{Q}}_i = R_{i,k} \bm{Q}_i, \quad \tilde{\bm{K}}_i = R_{i,k} \bm{K}_i,\quad
% \end{equation}
% \begin{equation}
\text{RoPE Attention}(X) = \text{Softmax} \left( \frac{\tilde{\bm{Q}}\tilde{\bm{K}}^T}{\sqrt{d_k}} \right) V\,.
\end{equation}
Notice that any computation in this layer which involves more than one tensor, is a matrix multiplication function. Adopting the existing baseline, we use the uniform relevance propagation rule, distributing the relevancy evenly between components. Thus, the relevancy scores of $Q,K,V,P$, with $P$ denoting the rotation matrix, is equal, and added together to the relevancy of the attention layer's output. The absorbing sink mechanism results in the following:
\begin{equation} \mathcal{R}^{(0)}= \mathcal{R}_{E},\quad \forall l \in [M]: \quad \mathcal{R}^{(l)} = \mathcal{R}^{(l-1)} + \mathcal{R}_{P}^{(l)} \end{equation}
Following this recursion we would get the exact same result as Eq.~\ref{eq:termToProveLemma3.2}
}
\end{proof}
{
\renewcommand{\thesection}{\arabic{section}}
\setcounter{lemma}{2}%result in 3
\setcounter{section}{3}
\begin{lemma}\label{lem:lemma3Appendix}
\textit{For attention‐level PE transformers, current LRP attribution rules achieve low faithfulness, especially when considering positional features.}
\end{lemma}
}
\begin{proof}[Proof of Lemma~\ref{lem:lemma3Appendix}]
% Write your proof here
\noindent We define a basic learning problem which relies solely on positional features, proving that existing LRP-based explanation methods which don't propagate relevance through positional encodings, will not produce faithful explanations.
% \myparagraph{Problem Settings.}
Let us assume we use an auto-regressive transformer model (e.g GPT), with a single causal self-attention with Alibi PE, and the Value projection replaced by an affine transformation (instead of a linear layer). Also, for brevity, let us consider scalar input tokens with sequence length of $L=2$, denoted by $x_{1}, x_{2}$. The final model uses the following keys ($K$), queries ($Q$), and values ($V$):
\begin{equation}
% Q_2 = W_Q X_2,\quad ,K_2 = W_K X_2,\quad K_1 = W_K X_1, \quad V_1 = W_V X_1,
\forall i \in [1,2]: \quad Q_i = W_Q X_i,\quad ,K_i = W_K X_i,\quad V_i = W_V X_i + b
\end{equation}
We apply the Alibi self-attention mechanism, and obtain the final output $O = (O_1, O_2)$:
\begin{equation}
A(i,j) = \frac{Q_{i}K_{j}}{\sqrt{d}} +m_{h}(i-j),\quad m_{h} = 1, \quad%
O_2 = A_{2,1} V_1 + A_{2,2} V_2
\end{equation}
To prevent the semantic representation from affecting the prediction, an optimal solution to this problem will assign zeros to $W_{Q}, W_{K}$, namely: $Q=K= \begin{pmatrix}
0 \\ 0
\end{pmatrix}$. For the Value projection, we assume: $V=W_{v}+b, \text{with } W_{v} = \begin{pmatrix}
0 \\ 0
\end{pmatrix}, b\neq\begin{pmatrix}
0 \\ 0
\end{pmatrix}$.
\myparagraph{Relevance propagation.}
Following our settings, we get:
\begin{equation}
A = \begin{pmatrix}
0 & 0 \\ 1 & 0
\end{pmatrix},\text{ giving us:}\quad Attention(Q,K,V)= A \times V = \begin{pmatrix}
0 & 0 \\ 1 & 0
\end{pmatrix}b
\end{equation}
For the backwards relevancy propagation, the relevancy scores of $Attention$ are distrusted between $A$ and $V$ based on the standard $\text{Gradient} \times \text{Input}$. Regardless, we now consider how relevancy scores of both terms $\mathcal{R}_{V},\mathcal{R}_{score}$ are propagated back to the input ${x}$.
\begin{itemize}
\item \textbf{$\mathcal{R}_{V}\rightarrow \mathcal{R}_{x}.\quad$} recall that $W_{v}$ are assigned with zeros. Given that the fundamental $\epsilon$-LRP rule for affine transformations ignores the bias term completely and uses the weights $W_{v}$ as a measure of weighting the relevancy scores, we get that zero relevancy scores are produced for both tokens.
\item \textbf{$\mathcal{R}_{A}\rightarrow \mathcal{R}_{x}.\quad$} Following the standard LRP rules, the positional terms would be considered a constant, and therefor, 100\% of the distribution would be directed to the queries and keys. Given that $W_{Q}, W_{K}$ are assigned with zeros, we again get zero relevancy scores being propagated to $x$.
\end{itemize}
Given that the relevancy scores propagated back from the attention layer are all assigned with zeros, we will get a final attribution map of zeroes, indicating the same level of impact for all tokens. This of course, yields an unfaithful explanation. In contrast, our method makes positional terms attributable, maintaining relevancy scores that would otherwise be zeroed out due to existing rules.
\end{proof}
\clearpage
%\renewcommand{\thesection}{\arabic{section}}
\renewcommand{\thesection}{\Alph{section}}
\setcounter{section}{5}
\section{Conservation Percentage Results \label{app:conservationPercentageResults}}
{
We measure the sum of relevance for DeiT model at different capacities: Tiny, Small, and Base. The figure provides clear visualization for the violation of the conservation property, with PE relevancy constituting 16.75\%, 22.39\%, and 9.22\% out of the total relevancy for Tiny, Small, and Base DeiT models, respectively.
% \begin{wrapfigure}{r}{0.46\textwidth}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\linewidth]{figures/vision/conservation_results.png}
\vspace{-4pt}
\caption{\small We assess both the positional relevance and the non-positional relevance for DeiT models at different capacities, visualizing the violation of conservation rule, with high non-negligible ratio between the entire relevance in the models 'ours' and positional-associated relevance 'PE Only'.}
\label{fig:conservationVisFig}
%\vspace{-35pt} no need vspace in appendix
\end{figure}
% \section{Experiments Compute Resources \label{app:computeResource}}
% All experiments were conducted using the PyTorch framework on publicly available datasets and executed on a single NVIDIA H100-80GB HBM3 GPU, running for at most 12 hours.
% }
% \section{Broader Impacts\label{app:impact}}
% Our work proposes a novel XAI technique that enhances the explainability of transformer models. XAI plays a critical role in ensuring the safe and responsible deployment of machine learning systems, particularly in high-stakes domains such as healthcare, finance, and law. It helps users and researchers understand, trust, and effectively audit model decisions. In particular, XAI facilitates the detection of biases, identification of failure modes, and debugging of unintended behaviors. By providing more accurate and faithful explanations for transformer-based models, our method contributes to greater transparency and accountability in DL systems. For all of those reasons, we believe this work will positively impact both the research community and practical applications by enabling safer and more interpretable use of powerful language models.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\end{document}
% \subsection{Ablations and Additional Analysis\label{sec:resultsAblations}}
% {\color{red}
% I this section...
% \noindent\textbf{Ablations.}\quad
% %Including Embedding Matrix with reform, Normalizations,
% }
% \noindent\textbf{Analyzed Conservation Property.\quad}
% \myparagraph{Analyzed Conservation Property.}
% We analyze the conservation property of our method compared to the AttnLRP baseline in Fig.~\ref{fig:conservation}. This experiment is conducted using three models: (i) LLaMA-2 7B, (ii) LLaMA-2 7B Quantized, and (iii) Tiny-LLaMA. All these models employ rotary PE and are fine-tuned on the IMDB dataset, where we measure relevance at the first, intermediate, and penultimate layers of the model. Specifically, we report how the relevance is distributed between the PEs, represented in orange, and the non-positional components, represented in blue. %The first component is unique to our method, while the 2nd is captured both by our method and AttnLRP.
% As our method allows explicit attribution to PE, this experiment quantifies the extent to which the lack of conservation in the LRP baseline~\citep{achtibat2024attnlrp} is recovered through our PA-LRP rules. % (defined in Sec.~\ref{sec:method})
% Our of 100\% of the attribution that exists in the top layer, a significant part is already lost in the penultimate layer. By the first layer, more than half of the relevancy is lost when reaching the first layer. The PE-only relevancy that our method adds on top of the relevancy that is retrieved by AttnLRP is able to recover much of the missing attribution. However, some relevance is still lost, e.g., due to the inability to attribute bias terms. %e Our method (which combines both types of relevance) better preserves the conservation property, highlighting that positional relevance, which was missing in previous work, is a major factor contributing to the lack of conservation in those methods.
% \begin{table}[h]\centering
% \small
% \setlength{\tabcolsep}{3pt}
% \begin{tabular}{c c c c c c }
% \toprule
% \multicolumn{1}{c}{Model} & \multicolumn{1}{c}{Method} & \multicolumn{2}{c}{Generation} & \multicolumn{2}{c}{Pruning} \\
% & & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$} & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$}\\
% \midrule
% % \multicolumn{5}{c}{LLaMa 2-7b}\\
% %
% LMA3 & AttnLRP & 0.037 & 152.934 & 0.037 & 152.808 \\
% LMA3 & PE Only & 0.314 & 41.680 & 0.304 & 43.037 \\
% LMA3 & Ours & \textbf{0.329} & \textbf{41.228} & \textbf{0.321} & \textbf{42.370} \\
% \bottomrule\\
% \end{tabular}
% \caption{\textbf{Perturbation Tests in NLP (Zero-Shot Models, Classification Regime). } \color{blue} Evaluation of LLaMa-3 8B ('LMA3') on pruning and generation perturbation tasks for multiple-choice question answering in zero-shot settings, using the ARC dataset. "AttnLRP" refers to the LRP baseline~\citep{achtibat2024attnlrp}. The metrics used are AUAC (area under activation curve, higher is better) and AU-MSE (area under the MSE, lower is better).\label{tab:pertubatoinZeroShot}}
% \end{table}
#########################
% \begin{table}[h]\centering
% \small
% \setlength{\tabcolsep}{3pt}
% \begin{tabular}{c c c c c c }
% \toprule
% \multicolumn{1}{c}{Model} & \multicolumn{1}{c}{Method} & \multicolumn{2}{c}{Generation} & \multicolumn{2}{c}{Pruning} \\
% & & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$} & \multicolumn{1}{c}{AUAC $\uparrow$} & \multicolumn{1}{c}{AU-MSE $\downarrow$}\\
% \midrule
% % \multicolumn{5}{c}{LLaMa 2-7b}\\
% %
% LMA3 & AttnLRP & 0.185 & 100.412 & 0.185 & 100.532 \\
% LMA3 & PE Only & 0.499 & 28.855 & 0.499 & 28.988 \\
% LMA3 & Ours & \textbf{0.506} & \textbf{28.660} & \textbf{0.505} & \textbf{28.795} \\
% \bottomrule\\
% \end{tabular}
% \caption{\textbf{Perturbation Tests in NLP (Zero-Shot Models, Next Token Prediction Regime).} \color{blue} Evaluation of LLaMa-3 8B ('LMA3') on pruning and generation perturbation tasks for next-word prediction, using the Wikipedia dataset. "AttnLRP" refers to the LRP baseline~\citep{achtibat2024attnlrp}. The metrics used are AUAC (area under activation curve, higher is better) and AU-MSE (area under the MSE, lower is better).\label{tab:pertubatoinWikiNLp}}
% \end{table}
% \noindent\textbf{Segmentation Tests in Vision.\quad}%
% As for segmentation tests, the empirical analysis in Tab.~\ref{tab:visionSegmentation} clearly demonstrates that our method outperform %s both
% the AttnLRP~\citep{achtibat2024attnlrp} baseline %
% %and the ablation variant, which focuses solely on positional maps, across all metrics and model sizes.
% In particular, our method improves over the baseline by 1\% points in Pixel Accuracy, 1.9\% points in mAP 1.9\% , and 2\% points in mIoU (Mean Intersection over Union). These results further highlight the importance of positionally associated relevance in effectively capturing spatial relationships and representing entire objects more accurately.
% \begin{table}[t]\centering
% % \small
% \setlength{\tabcolsep}{3pt}
% \begin{tabular}{c c c c c c c}
% \toprule
% M. Size & Method & Pixel Accuracy $\uparrow$ & mAP $\uparrow$ & mIoU $\uparrow$\\
% \midrule
% % & DeIT-Base & \\
% Base & AttnLRP~\citep{achtibat2024attnlrp} & 72.204 & 75.680 & 50.100 \\
% % Base & PE Only & 66.717 & 72.030 & 43.990 \\
% % Base & +PE & 66.882 & 71.130 & 43.330 \\
% Base & Ours & \textbf{72.698} & \textbf{77.260} & \textbf{51.400} \\
% \midrule
% % & DeIT7 Small & \\
% Small & AttnLRP~\citep{achtibat2024attnlrp} & 72.114 & 75.780 & 50.000 \\
% % Small &PE Only & 66.973 & 72.570 & 44.640 \\
% % Small &+PE & 67.213 & 71.010 & 43.520 \\
% Small & Ours & \textbf{73.060} & \textbf{77.340} & \textbf{51.700} \\
% \midrule
% % & DeIT-Tiny & \\
% % Tiny & Attibution~\citep{chefer2021transformer} & TBD & TBD & TBD \\
% % Tiny & Rollout~\citep{abnar2020quantifying} & TBD & TBD & TBD \\
% Tiny & AttnLRP~\citep{achtibat2024attnlrp} & 74.815 & 76.340 & 52.850 \\
% % Tiny & PE Only & 70.632 & 74.260 & 48.600 \\
% % Tiny &+PE & 69.269 & 71.360 & 45.370 \\
% Tiny & Ours & \textbf{76.613} & \textbf{78.920} & \textbf{55.920} \\
% \bottomrule
% \end{tabular}
% \caption{Segmentation performance of three DeiT variants across different model sizes on the ImageNet segmentation dataset~\citep{guillaumin2014imagenet}. 'M. Size' for models size. Higher is better.}
% %ALSO THE NAMING + PE IS SOME SORT OF AN ABLATION FOR US . I SUGGEST ``Baseline LRP'', ``XXX Ablation'', ``PA-LRP''. ALSO SINCE THERE ARE TWO CONTRIBUTIONS, SHOULD NOT WE HAVE TWO ABLATIONS}
% \label{tab:visionSegmentation}
% \end{table}
% \begin{figure*}[h]
% \centering
% \input{figures/NLP/samples/2/heatmap_9_baseline}\\
% \input{figures/NLP/samples/2/heatmap_9_PE} \\
% \input{figures/NLP/samples/2/heatmap_9_LRP+PE+REFORM}\\
% \caption{Baseline, PE, Ours}
% \label{fig:colored-text}
% \end{figure*}
\begin{figure*}[t]
\centering \small
\resizebox{0.89\textwidth}{!}{%
\begin{tabular}{l}
\input{figures/NLP/samples/3/heatmap_11_baseline}\\
\input{figures/NLP/samples/3/heatmap_11_PE}\\
\input{figures/NLP/samples/3/heatmap_11_LRP+PE+REFORM}\\
\multicolumn{1}{c}{(a)} \\
% \begin{figure*}[h]
\input{figures/NLP/samples/7/heatmap_25_baseline}\\
\input{figures/NLP/samples/7/heatmap_25_PE} \\
\input{figures/NLP/samples/7/heatmap_25_PE+LRP}\\
\multicolumn{1}{c}{(b)} \\
\end{tabular}
}
\vspace{-5pt}
\caption{\small \textbf{Qualitative Results in NLP.} Both groups (a) and (b) present results from different explanation methods for the same example obtained from the IMDB benchmark. In each group, the first row represents the AttnLRP baseline, followed by the PE-only variant in the middle, and finally, our maps at the end.\label{fig:qualitative_cmp_NLP}}
\end{figure*}
% \begin{figure*}[h]
% \centering
% \input{figures/NLP/samples/4/heatmap_13_baseline}\\
% \input{figures/NLP/samples/4/heatmap_13_PE}\\
% \input{figures/NLP/samples/4/heatmap_13_LRP+PE+REFORM} \\
% \caption{Baseline, PE, Ours}
% \label{fig:colored-text}
% \end{figure*}
% \begin{figure*}[h]
% \centering
% \input{figures/NLP/samples/5/heatmap_18_baseline}\\
% \input{figures/NLP/samples/5/heatmap_18_PE}\\
% \input{figures/NLP/samples/5/heatmap_18_PE+LRP} \\
% \caption{Baseline, PE, Ours}
% \label{fig:colored-text}
% \end{figure*}
%\newpage
%\begin{figure*}[t]
%\centering \small
%\resizebox{0.89\textwidth}{!}{%
% \begin{tabular}{l}
%
% %\input{figures/NLP/samples/arc_examples/arc_heat%map_baseline}\\
% %\input{figures/NLP/samples/arc_examples/arc_heat%map_ours} \\
% \end{tabular}
%}
% \vspace{-6pt}
%\caption{\textbf{Qualitative Results for MCQA.} Binary visualization for the top-70\% tokens with highest relevancy, with the first row employing AttnLRP baseline, and second row employing our method. During the 3rd/7th perturbation steps for pruning/generation scenarios respectively, the model makes a prediction solely based on the words marked in red. Critical words such as "decrease" and "nut", which are marked only by employing our method, exemplifies how the model's and even a human understanding could be drastically altered based on if these tokens appear or not.\label{fig:qualitative_cmp_NLP_ARC}}
%\end{figure*}
\newpage
% \begin{figure}
% \centering
% %\begin{tabular}{cccc}
% % \begin{tabular}{@{}c@{}c@{}c@{}c@{}}
% % \begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
% \begin{tabular}{@{\hskip 0.03in}c@{\hskip 0.03in}c@{\hskip 0.03in}}
% %Input & PA-LRP (Ours) &
% Positional & Semanitcal \\
% \cmidrule(lr){1-2}
% %\midrule
% % \includegraphics[width=0.122\textwidth]{figures/vision/img1/img_196.png} &
% % \includegraphics[width=0.122\textwidth]{figures/vision/img1/196_basic_custom_lrp.png} &
% \includegraphics[width=0.122\textwidth]{figures/vision/img1/196_basic_custom_lrp_PE_ONLY.png} & \includegraphics[width=0.122\textwidth]{figures/vision/img1/196_basic_custom_lrp_SEMANTIC_ONLY.png}
% \\
% % \includegraphics[width=0.122\textwidth]{figures/vision/img5/img_165.png} & \includegraphics[width=0.122\textwidth]{figures/vision/img5/165_basic_custom_lrp.png} &
% \includegraphics[width=0.122\textwidth]{figures/vision/img5/165_basic_custom_lrp_PE_ONLY.png} &
% \includegraphics[width=0.122\textwidth]{figures/vision/img5/165_basic_custom_lrp_SEMANTIC_ONLY.png}\\
% % \\
% % \includegraphics[width=0.122\textwidth]{figures/vision/img7/img_92.png} &
% % \includegraphics[width=0.122\textwidth]{figures/vision/img7/92_basic_custom_lrp.png} &
% \includegraphics[width=0.122\textwidth]{figures/vision/img7/92_basic_custom_lrp_PE_ONLY.png} & \includegraphics[width=0.122\textwidth]{figures/vision/img7/92_basic_custom_lrp_SEMANTIC_ONLY.png}
% \\
% % \includegraphics[width=0.122\textwidth]{figures/vision/img6/img_157.png} & \includegraphics[width=0.122\textwidth]{figures/vision/img6/157_basic_custom_lrp.png} &
% \includegraphics[width=0.122\textwidth]{figures/vision/img6/157_basic_custom_lrp_PE_ONLY.png} &
% \includegraphics[width=0.122\textwidth]
% {figures/vision/img6/157_basic_custom_lrp_SEMANTIC_ONLY.png}
% \end{tabular}
% \caption{{\color{red}PLACEHOLDER}\textbf{HTBD:} TBD}\label{fig:qualitative_cmp_vision}
% \end{figure}
\fbox{
\parbox{\textwidth}{
\setlength\fboxsep{0pt}
\colorbox[RGB]{255,0,0}{\strut <s>} \colorbox[RGB]{255,226,226}{\strut Great} \colorbox[RGB]{255,248,248}{\strut just} \colorbox[RGB]{255,234,234}{\strut great} \colorbox[RGB]{255,222,222}{\strut !} \colorbox[RGB]{246,246,255}{\strut The} \colorbox[RGB]{131,131,255}{\strut West} \colorbox[RGB]{255,252,252}{\strut Coast} \colorbox[RGB]{252,252,255}{\strut got} \colorbox[RGB]{255,254,254}{\strut "} \colorbox[RGB]{255,254,254}{\strut Dir} \colorbox[RGB]{255,254,254}{\strut ty} \colorbox[RGB]{254,254,255}{\strut "} \colorbox[RGB]{243,243,255}{\strut Harry} \colorbox[RGB]{255,254,254}{\strut Cal} \colorbox[RGB]{255,252,252}{\strut la} \colorbox[RGB]{255,250,250}{\strut han} \colorbox[RGB]{255,254,254}{\strut ,} \colorbox[RGB]{254,254,255}{\strut the} \colorbox[RGB]{250,250,255}{\strut East} \colorbox[RGB]{255,254,254}{\strut Coast} \colorbox[RGB]{254,254,255}{\strut got} \colorbox[RGB]{255,242,242}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,252,252}{\strut y} \colorbox[RGB]{255,252,252}{\strut .} \colorbox[RGB]{255,252,252}{\strut B} \colorbox[RGB]{255,254,254}{\strut urt} \colorbox[RGB]{227,227,255}{\strut Reyn} \colorbox[RGB]{246,246,255}{\strut olds} \colorbox[RGB]{250,250,255}{\strut plays} \colorbox[RGB]{255,250,250}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,252,252}{\strut y} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{188,188,255}{\strut "} \colorbox[RGB]{255,250,250}{\strut Sh} \colorbox[RGB]{254,254,255}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,244,244}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,248,248}{\strut Machine} \colorbox[RGB]{252,252,255}{\strut "} \colorbox[RGB]{255,248,248}{\strut and} \colorbox[RGB]{255,242,242}{\strut I} \colorbox[RGB]{255,240,240}{\strut enjoyed} \colorbox[RGB]{255,226,226}{\strut every} \colorbox[RGB]{255,250,250}{\strut minute} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,248,248}{\strut it} \colorbox[RGB]{255,226,226}{\strut .} \colorbox[RGB]{255,246,246}{\strut Play} \colorbox[RGB]{255,254,254}{\strut ing} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{246,246,255}{\strut ma} \colorbox[RGB]{255,254,254}{\strut ver} \colorbox[RGB]{255,254,254}{\strut ick} \colorbox[RGB]{255,254,254}{\strut n} \colorbox[RGB]{255,254,254}{\strut arc} \colorbox[RGB]{255,254,254}{\strut ot} \colorbox[RGB]{254,254,255}{\strut ics} \colorbox[RGB]{255,252,252}{\strut cop} \colorbox[RGB]{255,250,250}{\strut in} \colorbox[RGB]{255,248,248}{\strut Atlanta} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,252,252}{\strut G} \colorbox[RGB]{255,250,250}{\strut A} \colorbox[RGB]{252,252,255}{\strut is} \colorbox[RGB]{255,252,252}{\strut just} \colorbox[RGB]{254,254,255}{\strut what} \colorbox[RGB]{255,250,250}{\strut everyone} \colorbox[RGB]{255,252,252}{\strut wants} \colorbox[RGB]{255,244,244}{\strut .} \colorbox[RGB]{243,243,255}{\strut Instead} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,254,254}{\strut susp} \colorbox[RGB]{255,254,254}{\strut ension} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{248,248,255}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{254,254,255}{\strut sent} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,250,250}{\strut vice} \colorbox[RGB]{254,254,255}{\strut squad} \colorbox[RGB]{254,254,255}{\strut .} \colorbox[RGB]{242,242,255}{\strut Like} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut D} \colorbox[RGB]{254,254,255}{\strut irty} \colorbox[RGB]{252,252,255}{\strut Harry} \colorbox[RGB]{254,254,255}{\strut mov} \colorbox[RGB]{255,254,254}{\strut ies} \colorbox[RGB]{250,250,255}{\strut or} \colorbox[RGB]{255,254,254}{\strut any} \colorbox[RGB]{255,254,254}{\strut other} \colorbox[RGB]{255,246,246}{\strut cop} \colorbox[RGB]{254,254,255}{\strut mov} \colorbox[RGB]{255,254,254}{\strut ies} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,252,252}{\strut captain} \colorbox[RGB]{254,254,255}{\strut is} \colorbox[RGB]{254,254,255}{\strut always} \colorbox[RGB]{254,254,255}{\strut going} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{254,254,255}{\strut be} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{252,252,255}{\strut j} \colorbox[RGB]{248,248,255}{\strut erk} \colorbox[RGB]{246,246,255}{\strut .} \colorbox[RGB]{255,254,254}{\strut When} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,254,254}{\strut kid} \colorbox[RGB]{255,254,254}{\strut ,} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,254,254}{\strut curious} \colorbox[RGB]{255,252,252}{\strut what} \colorbox[RGB]{255,252,252}{\strut that} \colorbox[RGB]{252,252,255}{\strut movie} \colorbox[RGB]{252,252,255}{\strut meant} \colorbox[RGB]{252,252,255}{\strut "} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{254,254,255}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,254,254}{\strut Machine} \colorbox[RGB]{255,254,254}{\strut ".} \colorbox[RGB]{255,248,248}{\strut Well} \colorbox[RGB]{254,254,255}{\strut I} \colorbox[RGB]{255,252,252}{\strut knew} \colorbox[RGB]{254,254,255}{\strut who} \colorbox[RGB]{254,254,255}{\strut played} \colorbox[RGB]{255,254,254}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,254,254}{\strut ,} \colorbox[RGB]{254,254,255}{\strut I} \colorbox[RGB]{254,254,255}{\strut wonder} \colorbox[RGB]{254,254,255}{\strut what} \colorbox[RGB]{254,254,255}{\strut his} \colorbox[RGB]{255,254,254}{\strut machine} \colorbox[RGB]{254,254,255}{\strut was} \colorbox[RGB]{252,252,255}{\strut .} \colorbox[RGB]{255,254,254}{\strut It} \colorbox[RGB]{254,254,255}{\strut was} \colorbox[RGB]{255,254,254}{\strut his} \colorbox[RGB]{254,254,255}{\strut GROUP} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,252,252}{\strut fellow} \colorbox[RGB]{255,254,254}{\strut c} \colorbox[RGB]{255,252,252}{\strut ops} \colorbox[RGB]{255,254,254}{\strut .} \colorbox[RGB]{248,248,255}{\strut After} \colorbox[RGB]{255,254,254}{\strut un} \colorbox[RGB]{255,254,254}{\strut cover} \colorbox[RGB]{255,254,254}{\strut ing} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,250,250}{\strut murder} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,252,252}{\strut he} \colorbox[RGB]{254,254,255}{\strut goes} \colorbox[RGB]{255,254,254}{\strut all} \colorbox[RGB]{255,252,252}{\strut out} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut find} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut per} \colorbox[RGB]{254,254,255}{\strut p} \colorbox[RGB]{254,254,255}{\strut .} \colorbox[RGB]{252,252,255}{\strut When} \colorbox[RGB]{254,254,255}{\strut it} \colorbox[RGB]{250,250,255}{\strut turns} \colorbox[RGB]{254,254,255}{\strut out} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut big} \colorbox[RGB]{255,252,252}{\strut time} \colorbox[RGB]{255,248,248}{\strut mob} \colorbox[RGB]{255,254,254}{\strut b} \colorbox[RGB]{254,254,255}{\strut oss} \colorbox[RGB]{255,254,254}{\strut ,} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{252,252,255}{\strut doesn} \colorbox[RGB]{240,240,255}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,254,254}{\strut play} \colorbox[RGB]{255,254,254}{\strut around} \colorbox[RGB]{255,254,254}{\strut .} \colorbox[RGB]{254,254,255}{\strut When} \colorbox[RGB]{255,254,254}{\strut he} \colorbox[RGB]{254,254,255}{\strut gets} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut other} \colorbox[RGB]{248,248,255}{\strut prost} \colorbox[RGB]{255,254,254}{\strut itute} \colorbox[RGB]{255,252,252}{\strut into} \colorbox[RGB]{255,252,252}{\strut safety} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,254,254}{\strut f} \colorbox[RGB]{255,254,254}{\strut ights} \colorbox[RGB]{255,252,252}{\strut back} \colorbox[RGB]{255,252,252}{\strut hard} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{255,254,254}{\strut good} \colorbox[RGB]{246,246,255}{\strut despite} \colorbox[RGB]{254,254,255}{\strut losing} \colorbox[RGB]{254,254,255}{\strut a} \colorbox[RGB]{250,250,255}{\strut finger} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut th} \colorbox[RGB]{254,254,255}{\strut ug} \colorbox[RGB]{252,252,255}{\strut .} \colorbox[RGB]{242,242,255}{\strut And} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,244,244}{\strut also} \colorbox[RGB]{248,248,255}{\strut like} \colorbox[RGB]{254,254,255}{\strut the} \colorbox[RGB]{255,254,254}{\strut part} \colorbox[RGB]{255,254,254}{\strut where} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{252,252,255}{\strut bad} \colorbox[RGB]{252,252,255}{\strut gets} \colorbox[RGB]{252,252,255}{\strut blow} \colorbox[RGB]{254,254,255}{\strut n} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut building} \colorbox[RGB]{255,252,252}{\strut through} \colorbox[RGB]{254,254,255}{\strut a} \colorbox[RGB]{254,254,255}{\strut plate} \colorbox[RGB]{254,254,255}{\strut glass} \colorbox[RGB]{254,254,255}{\strut window} \colorbox[RGB]{242,242,255}{\strut .} \colorbox[RGB]{250,250,255}{\strut That} \colorbox[RGB]{248,248,255}{\strut was} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut B} \colorbox[RGB]{254,254,255}{\strut OM} \colorbox[RGB]{254,254,255}{\strut B} \colorbox[RGB]{255,246,246}{\strut !} \colorbox[RGB]{255,236,236}{\strut R} \colorbox[RGB]{255,238,238}{\strut andy} \colorbox[RGB]{254,254,255}{\strut C} \colorbox[RGB]{255,254,254}{\strut raw} \colorbox[RGB]{255,246,246}{\strut ford} \colorbox[RGB]{240,240,255}{\strut '} \colorbox[RGB]{255,250,250}{\strut s} \colorbox[RGB]{246,246,255}{\strut "} \colorbox[RGB]{255,250,250}{\strut St} \colorbox[RGB]{255,232,232}{\strut reet} \colorbox[RGB]{255,254,254}{\strut Life} \colorbox[RGB]{255,246,246}{\strut "} \colorbox[RGB]{255,230,230}{\strut really} \colorbox[RGB]{248,248,255}{\strut put} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{227,227,255}{\strut movie} \colorbox[RGB]{254,254,255}{\strut in} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,250,250}{\strut right} \colorbox[RGB]{252,252,255}{\strut m} \colorbox[RGB]{250,250,255}{\strut ood} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,244,244}{\strut and} \colorbox[RGB]{254,254,255}{\strut the} \colorbox[RGB]{250,250,255}{\strut movie} \colorbox[RGB]{255,244,244}{\strut itself} \colorbox[RGB]{255,252,252}{\strut is} \colorbox[RGB]{255,234,234}{\strut really} \colorbox[RGB]{255,246,246}{\strut a} \colorbox[RGB]{255,206,206}{\strut great} \colorbox[RGB]{255,240,240}{\strut hit} \colorbox[RGB]{255,240,240}{\strut to} \colorbox[RGB]{255,234,234}{\strut me} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{254,254,255}{\strut AL} \colorbox[RGB]{255,244,244}{\strut WA} \colorbox[RGB]{255,226,226}{\strut YS} \colorbox[RGB]{255,208,208}{\strut !} \colorbox[RGB]{250,250,255}{\strut R} \colorbox[RGB]{255,214,214}{\strut ating} \colorbox[RGB]{238,238,255}{\strut } \colorbox[RGB]{255,214,214}{\strut 4} \colorbox[RGB]{255,242,242}{\strut out} \colorbox[RGB]{255,238,238}{\strut of} \colorbox[RGB]{255,234,234}{\strut } \colorbox[RGB]{255,162,162}{\strut 5} \colorbox[RGB]{255,178,178}{\strut stars} \colorbox[RGB]{255,33,33}{\strut .}}}
\fbox{
\parbox{\textwidth}{
\setlength\fboxsep{0pt}
\colorbox[RGB]{255,0,0}{\strut <s>} \colorbox[RGB]{255,226,226}{\strut Great} \colorbox[RGB]{255,246,246}{\strut just} \colorbox[RGB]{255,234,234}{\strut great} \colorbox[RGB]{255,226,226}{\strut !} \colorbox[RGB]{255,252,252}{\strut The} \colorbox[RGB]{240,240,255}{\strut West} \colorbox[RGB]{255,250,250}{\strut Coast} \colorbox[RGB]{255,252,252}{\strut got} \colorbox[RGB]{255,252,252}{\strut "} \colorbox[RGB]{255,252,252}{\strut Dir} \colorbox[RGB]{255,252,252}{\strut ty} \colorbox[RGB]{255,252,252}{\strut "} \colorbox[RGB]{255,254,254}{\strut Harry} \colorbox[RGB]{255,250,250}{\strut Cal} \colorbox[RGB]{255,254,254}{\strut la} \colorbox[RGB]{255,246,246}{\strut han} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut East} \colorbox[RGB]{255,252,252}{\strut Coast} \colorbox[RGB]{255,252,252}{\strut got} \colorbox[RGB]{255,250,250}{\strut Sh} \colorbox[RGB]{255,252,252}{\strut ark} \colorbox[RGB]{255,250,250}{\strut y} \colorbox[RGB]{255,244,244}{\strut .} \colorbox[RGB]{255,250,250}{\strut B} \colorbox[RGB]{255,250,250}{\strut urt} \colorbox[RGB]{254,254,255}{\strut Reyn} \colorbox[RGB]{255,252,252}{\strut olds} \colorbox[RGB]{255,250,250}{\strut plays} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,250,250}{\strut in} \colorbox[RGB]{248,248,255}{\strut "} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,252,252}{\strut y} \colorbox[RGB]{255,248,248}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,250,250}{\strut Machine} \colorbox[RGB]{255,252,252}{\strut "} \colorbox[RGB]{255,250,250}{\strut and} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,240,240}{\strut enjoyed} \colorbox[RGB]{255,236,236}{\strut every} \colorbox[RGB]{255,248,248}{\strut minute} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,248,248}{\strut it} \colorbox[RGB]{255,230,230}{\strut .} \colorbox[RGB]{255,246,246}{\strut Play} \colorbox[RGB]{255,252,252}{\strut ing} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{254,254,255}{\strut ma} \colorbox[RGB]{255,254,254}{\strut ver} \colorbox[RGB]{255,254,254}{\strut ick} \colorbox[RGB]{255,254,254}{\strut n} \colorbox[RGB]{255,254,254}{\strut arc} \colorbox[RGB]{255,252,252}{\strut ot} \colorbox[RGB]{255,254,254}{\strut ics} \colorbox[RGB]{255,250,250}{\strut cop} \colorbox[RGB]{255,250,250}{\strut in} \colorbox[RGB]{255,248,248}{\strut Atlanta} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,250,250}{\strut G} \colorbox[RGB]{255,248,248}{\strut A} \colorbox[RGB]{255,252,252}{\strut is} \colorbox[RGB]{255,252,252}{\strut just} \colorbox[RGB]{255,250,250}{\strut what} \colorbox[RGB]{255,250,250}{\strut everyone} \colorbox[RGB]{255,252,252}{\strut wants} \colorbox[RGB]{255,242,242}{\strut .} \colorbox[RGB]{254,254,255}{\strut Instead} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,254,254}{\strut susp} \colorbox[RGB]{255,254,254}{\strut ension} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,254,254}{\strut sent} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,252,252}{\strut vice} \colorbox[RGB]{255,254,254}{\strut squad} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{254,254,255}{\strut Like} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut D} \colorbox[RGB]{255,254,254}{\strut irty} \colorbox[RGB]{255,254,254}{\strut Harry} \colorbox[RGB]{255,254,254}{\strut mov} \colorbox[RGB]{255,252,252}{\strut ies} \colorbox[RGB]{255,252,252}{\strut or} \colorbox[RGB]{255,252,252}{\strut any} \colorbox[RGB]{255,254,254}{\strut other} \colorbox[RGB]{255,246,246}{\strut cop} \colorbox[RGB]{255,254,254}{\strut mov} \colorbox[RGB]{255,254,254}{\strut ies} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,252,252}{\strut captain} \colorbox[RGB]{255,254,254}{\strut is} \colorbox[RGB]{255,254,254}{\strut always} \colorbox[RGB]{255,254,254}{\strut going} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut j} \colorbox[RGB]{255,252,252}{\strut erk} \colorbox[RGB]{255,252,252}{\strut .} \colorbox[RGB]{255,252,252}{\strut When} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut kid} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,252,252}{\strut curious} \colorbox[RGB]{255,252,252}{\strut what} \colorbox[RGB]{255,252,252}{\strut that} \colorbox[RGB]{255,252,252}{\strut movie} \colorbox[RGB]{255,252,252}{\strut meant} \colorbox[RGB]{255,252,252}{\strut "} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,252,252}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,252,252}{\strut Machine} \colorbox[RGB]{255,246,246}{\strut ".} \colorbox[RGB]{255,250,250}{\strut Well} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,252,252}{\strut knew} \colorbox[RGB]{255,252,252}{\strut who} \colorbox[RGB]{255,254,254}{\strut played} \colorbox[RGB]{255,254,254}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,254,254}{\strut wonder} \colorbox[RGB]{255,254,254}{\strut what} \colorbox[RGB]{255,254,254}{\strut his} \colorbox[RGB]{255,252,252}{\strut machine} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,252,252}{\strut It} \colorbox[RGB]{255,252,252}{\strut was} \colorbox[RGB]{255,252,252}{\strut his} \colorbox[RGB]{255,250,250}{\strut GROUP} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,252,252}{\strut fellow} \colorbox[RGB]{255,254,254}{\strut c} \colorbox[RGB]{255,250,250}{\strut ops} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,254,254}{\strut After} \colorbox[RGB]{255,254,254}{\strut un} \colorbox[RGB]{255,254,254}{\strut cover} \colorbox[RGB]{255,252,252}{\strut ing} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,250,250}{\strut murder} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,252,252}{\strut he} \colorbox[RGB]{255,254,254}{\strut goes} \colorbox[RGB]{255,252,252}{\strut all} \colorbox[RGB]{255,252,252}{\strut out} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut find} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut per} \colorbox[RGB]{255,254,254}{\strut p} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,254,254}{\strut When} \colorbox[RGB]{255,254,254}{\strut it} \colorbox[RGB]{255,254,254}{\strut turns} \colorbox[RGB]{255,252,252}{\strut out} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut big} \colorbox[RGB]{255,252,252}{\strut time} \colorbox[RGB]{255,250,250}{\strut mob} \colorbox[RGB]{255,254,254}{\strut b} \colorbox[RGB]{255,254,254}{\strut oss} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,254,254}{\strut doesn} \colorbox[RGB]{255,248,248}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,254,254}{\strut play} \colorbox[RGB]{255,254,254}{\strut around} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,254,254}{\strut When} \colorbox[RGB]{255,254,254}{\strut he} \colorbox[RGB]{255,254,254}{\strut gets} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut other} \colorbox[RGB]{254,254,255}{\strut prost} \colorbox[RGB]{255,252,252}{\strut itute} \colorbox[RGB]{255,252,252}{\strut into} \colorbox[RGB]{255,252,252}{\strut safety} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,254,254}{\strut f} \colorbox[RGB]{255,252,252}{\strut ights} \colorbox[RGB]{255,252,252}{\strut back} \colorbox[RGB]{255,250,250}{\strut hard} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{255,252,252}{\strut good} \colorbox[RGB]{254,254,255}{\strut despite} \colorbox[RGB]{255,254,254}{\strut losing} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,254,254}{\strut finger} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut th} \colorbox[RGB]{255,254,254}{\strut ug} \colorbox[RGB]{255,244,244}{\strut .} \colorbox[RGB]{255,254,254}{\strut And} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{255,244,244}{\strut also} \colorbox[RGB]{255,248,248}{\strut like} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,250,250}{\strut part} \colorbox[RGB]{255,250,250}{\strut where} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut bad} \colorbox[RGB]{255,254,254}{\strut gets} \colorbox[RGB]{254,254,255}{\strut blow} \colorbox[RGB]{255,252,252}{\strut n} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut building} \colorbox[RGB]{255,252,252}{\strut through} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,254,254}{\strut plate} \colorbox[RGB]{255,254,254}{\strut glass} \colorbox[RGB]{255,254,254}{\strut window} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,252,252}{\strut That} \colorbox[RGB]{255,250,250}{\strut was} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,252,252}{\strut B} \colorbox[RGB]{255,252,252}{\strut OM} \colorbox[RGB]{255,250,250}{\strut B} \colorbox[RGB]{255,232,232}{\strut !} \colorbox[RGB]{255,240,240}{\strut R} \colorbox[RGB]{255,236,236}{\strut andy} \colorbox[RGB]{255,250,250}{\strut C} \colorbox[RGB]{255,252,252}{\strut raw} \colorbox[RGB]{255,230,230}{\strut ford} \colorbox[RGB]{255,248,248}{\strut '} \colorbox[RGB]{255,244,244}{\strut s} \colorbox[RGB]{255,248,248}{\strut "} \colorbox[RGB]{255,250,250}{\strut St} \colorbox[RGB]{255,242,242}{\strut reet} \colorbox[RGB]{255,250,250}{\strut Life} \colorbox[RGB]{255,240,240}{\strut "} \colorbox[RGB]{255,230,230}{\strut really} \colorbox[RGB]{255,246,246}{\strut put} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,254,254}{\strut movie} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,248,248}{\strut right} \colorbox[RGB]{255,254,254}{\strut m} \colorbox[RGB]{255,248,248}{\strut ood} \colorbox[RGB]{255,246,246}{\strut ,} \colorbox[RGB]{255,240,240}{\strut and} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,242,242}{\strut movie} \colorbox[RGB]{255,242,242}{\strut itself} \colorbox[RGB]{255,246,246}{\strut is} \colorbox[RGB]{255,242,242}{\strut really} \colorbox[RGB]{255,246,246}{\strut a} \colorbox[RGB]{255,214,214}{\strut great} \colorbox[RGB]{255,234,234}{\strut hit} \colorbox[RGB]{255,234,234}{\strut to} \colorbox[RGB]{255,224,224}{\strut me} \colorbox[RGB]{255,236,236}{\strut ,} \colorbox[RGB]{255,248,248}{\strut AL} \colorbox[RGB]{255,248,248}{\strut WA} \colorbox[RGB]{255,216,216}{\strut YS} \colorbox[RGB]{255,184,184}{\strut !} \colorbox[RGB]{255,238,238}{\strut R} \colorbox[RGB]{255,184,184}{\strut ating} \colorbox[RGB]{255,220,220}{\strut } \colorbox[RGB]{255,210,210}{\strut 4} \colorbox[RGB]{255,238,238}{\strut out} \colorbox[RGB]{255,224,224}{\strut of} \colorbox[RGB]{255,226,226}{\strut } \colorbox[RGB]{255,147,147}{\strut 5} \colorbox[RGB]{255,156,156}{\strut stars} \colorbox[RGB]{255,73,73}{\strut .}}}
\fbox{
\parbox{\textwidth}{
\setlength\fboxsep{0pt}
\colorbox[RGB]{255,36,36}{\strut <s>} \colorbox[RGB]{255,214,214}{\strut Great} \colorbox[RGB]{255,244,244}{\strut just} \colorbox[RGB]{255,226,226}{\strut great} \colorbox[RGB]{255,216,216}{\strut !} \colorbox[RGB]{255,250,250}{\strut The} \colorbox[RGB]{227,227,255}{\strut West} \colorbox[RGB]{255,248,248}{\strut Coast} \colorbox[RGB]{255,252,252}{\strut got} \colorbox[RGB]{255,250,250}{\strut "} \colorbox[RGB]{255,252,252}{\strut Dir} \colorbox[RGB]{255,252,252}{\strut ty} \colorbox[RGB]{255,250,250}{\strut "} \colorbox[RGB]{255,252,252}{\strut Harry} \colorbox[RGB]{255,250,250}{\strut Cal} \colorbox[RGB]{255,254,254}{\strut la} \colorbox[RGB]{255,246,246}{\strut han} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut East} \colorbox[RGB]{255,252,252}{\strut Coast} \colorbox[RGB]{255,250,250}{\strut got} \colorbox[RGB]{255,248,248}{\strut Sh} \colorbox[RGB]{255,252,252}{\strut ark} \colorbox[RGB]{255,250,250}{\strut y} \colorbox[RGB]{255,244,244}{\strut .} \colorbox[RGB]{255,250,250}{\strut B} \colorbox[RGB]{255,250,250}{\strut urt} \colorbox[RGB]{252,252,255}{\strut Reyn} \colorbox[RGB]{255,250,250}{\strut olds} \colorbox[RGB]{255,248,248}{\strut plays} \colorbox[RGB]{255,250,250}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,252,252}{\strut y} \colorbox[RGB]{255,248,248}{\strut in} \colorbox[RGB]{243,243,255}{\strut "} \colorbox[RGB]{255,250,250}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,252,252}{\strut y} \colorbox[RGB]{255,246,246}{\strut '} \colorbox[RGB]{255,252,252}{\strut s} \colorbox[RGB]{255,248,248}{\strut Machine} \colorbox[RGB]{255,250,250}{\strut "} \colorbox[RGB]{255,246,246}{\strut and} \colorbox[RGB]{255,242,242}{\strut I} \colorbox[RGB]{255,236,236}{\strut enjoyed} \colorbox[RGB]{255,232,232}{\strut every} \colorbox[RGB]{255,248,248}{\strut minute} \colorbox[RGB]{255,250,250}{\strut of} \colorbox[RGB]{255,248,248}{\strut it} \colorbox[RGB]{255,230,230}{\strut .} \colorbox[RGB]{255,244,244}{\strut Play} \colorbox[RGB]{255,250,250}{\strut ing} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{254,254,255}{\strut ma} \colorbox[RGB]{255,254,254}{\strut ver} \colorbox[RGB]{255,252,252}{\strut ick} \colorbox[RGB]{255,254,254}{\strut n} \colorbox[RGB]{255,254,254}{\strut arc} \colorbox[RGB]{255,252,252}{\strut ot} \colorbox[RGB]{255,254,254}{\strut ics} \colorbox[RGB]{255,250,250}{\strut cop} \colorbox[RGB]{255,250,250}{\strut in} \colorbox[RGB]{255,246,246}{\strut Atlanta} \colorbox[RGB]{255,246,246}{\strut ,} \colorbox[RGB]{255,250,250}{\strut G} \colorbox[RGB]{255,248,248}{\strut A} \colorbox[RGB]{255,250,250}{\strut is} \colorbox[RGB]{255,250,250}{\strut just} \colorbox[RGB]{255,250,250}{\strut what} \colorbox[RGB]{255,248,248}{\strut everyone} \colorbox[RGB]{255,250,250}{\strut wants} \colorbox[RGB]{255,242,242}{\strut .} \colorbox[RGB]{254,254,255}{\strut Instead} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,254,254}{\strut susp} \colorbox[RGB]{255,254,254}{\strut ension} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{255,252,252}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,254,254}{\strut sent} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,252,252}{\strut vice} \colorbox[RGB]{255,254,254}{\strut squad} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{254,254,255}{\strut Like} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut D} \colorbox[RGB]{255,254,254}{\strut irty} \colorbox[RGB]{255,252,252}{\strut Harry} \colorbox[RGB]{255,254,254}{\strut mov} \colorbox[RGB]{255,252,252}{\strut ies} \colorbox[RGB]{255,254,254}{\strut or} \colorbox[RGB]{255,252,252}{\strut any} \colorbox[RGB]{255,254,254}{\strut other} \colorbox[RGB]{255,246,246}{\strut cop} \colorbox[RGB]{255,254,254}{\strut mov} \colorbox[RGB]{255,254,254}{\strut ies} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,252,252}{\strut captain} \colorbox[RGB]{255,254,254}{\strut is} \colorbox[RGB]{255,254,254}{\strut always} \colorbox[RGB]{254,254,255}{\strut going} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut j} \colorbox[RGB]{255,254,254}{\strut erk} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,252,252}{\strut When} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut kid} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,252,252}{\strut curious} \colorbox[RGB]{255,252,252}{\strut what} \colorbox[RGB]{255,252,252}{\strut that} \colorbox[RGB]{255,252,252}{\strut movie} \colorbox[RGB]{255,254,254}{\strut meant} \colorbox[RGB]{255,250,250}{\strut "} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,252,252}{\strut y} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,252,252}{\strut Machine} \colorbox[RGB]{255,244,244}{\strut ".} \colorbox[RGB]{255,246,246}{\strut Well} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,252,252}{\strut knew} \colorbox[RGB]{255,254,254}{\strut who} \colorbox[RGB]{255,254,254}{\strut played} \colorbox[RGB]{255,254,254}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,254,254}{\strut wonder} \colorbox[RGB]{255,254,254}{\strut what} \colorbox[RGB]{255,254,254}{\strut his} \colorbox[RGB]{255,254,254}{\strut machine} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,252,252}{\strut It} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,252,252}{\strut his} \colorbox[RGB]{255,250,250}{\strut GROUP} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,252,252}{\strut fellow} \colorbox[RGB]{255,254,254}{\strut c} \colorbox[RGB]{255,252,252}{\strut ops} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,254,254}{\strut After} \colorbox[RGB]{255,254,254}{\strut un} \colorbox[RGB]{255,254,254}{\strut cover} \colorbox[RGB]{255,254,254}{\strut ing} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,250,250}{\strut murder} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,252,252}{\strut he} \colorbox[RGB]{255,254,254}{\strut goes} \colorbox[RGB]{255,252,252}{\strut all} \colorbox[RGB]{255,252,252}{\strut out} \colorbox[RGB]{255,252,252}{\strut to} \colorbox[RGB]{255,254,254}{\strut find} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut per} \colorbox[RGB]{255,254,254}{\strut p} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,254,254}{\strut When} \colorbox[RGB]{255,254,254}{\strut it} \colorbox[RGB]{254,254,255}{\strut turns} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut big} \colorbox[RGB]{255,252,252}{\strut time} \colorbox[RGB]{255,250,250}{\strut mob} \colorbox[RGB]{255,254,254}{\strut b} \colorbox[RGB]{255,254,254}{\strut oss} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,252,252}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,254,254}{\strut doesn} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,254,254}{\strut play} \colorbox[RGB]{255,254,254}{\strut around} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,252,252}{\strut When} \colorbox[RGB]{255,254,254}{\strut he} \colorbox[RGB]{255,254,254}{\strut gets} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut other} \colorbox[RGB]{254,254,255}{\strut prost} \colorbox[RGB]{255,254,254}{\strut itute} \colorbox[RGB]{255,252,252}{\strut into} \colorbox[RGB]{255,252,252}{\strut safety} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Sh} \colorbox[RGB]{255,254,254}{\strut ark} \colorbox[RGB]{255,254,254}{\strut y} \colorbox[RGB]{255,254,254}{\strut f} \colorbox[RGB]{255,252,252}{\strut ights} \colorbox[RGB]{255,252,252}{\strut back} \colorbox[RGB]{255,250,250}{\strut hard} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{255,252,252}{\strut good} \colorbox[RGB]{254,254,255}{\strut despite} \colorbox[RGB]{255,254,254}{\strut losing} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{254,254,255}{\strut finger} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut th} \colorbox[RGB]{255,254,254}{\strut ug} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,254,254}{\strut And} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{255,244,244}{\strut also} \colorbox[RGB]{255,250,250}{\strut like} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,252,252}{\strut part} \colorbox[RGB]{255,250,250}{\strut where} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut bad} \colorbox[RGB]{255,254,254}{\strut gets} \colorbox[RGB]{255,254,254}{\strut blow} \colorbox[RGB]{255,254,254}{\strut n} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut building} \colorbox[RGB]{255,252,252}{\strut through} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{254,254,255}{\strut plate} \colorbox[RGB]{255,254,254}{\strut glass} \colorbox[RGB]{255,254,254}{\strut window} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,254,254}{\strut That} \colorbox[RGB]{255,252,252}{\strut was} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut B} \colorbox[RGB]{255,254,254}{\strut OM} \colorbox[RGB]{255,250,250}{\strut B} \colorbox[RGB]{255,230,230}{\strut !} \colorbox[RGB]{255,236,236}{\strut R} \colorbox[RGB]{255,236,236}{\strut andy} \colorbox[RGB]{255,252,252}{\strut C} \colorbox[RGB]{255,252,252}{\strut raw} \colorbox[RGB]{255,238,238}{\strut ford} \colorbox[RGB]{255,248,248}{\strut '} \colorbox[RGB]{255,244,244}{\strut s} \colorbox[RGB]{255,242,242}{\strut "} \colorbox[RGB]{255,250,250}{\strut St} \colorbox[RGB]{255,240,240}{\strut reet} \colorbox[RGB]{255,250,250}{\strut Life} \colorbox[RGB]{255,240,240}{\strut "} \colorbox[RGB]{255,232,232}{\strut really} \colorbox[RGB]{255,250,250}{\strut put} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,248,248}{\strut movie} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,248,248}{\strut right} \colorbox[RGB]{255,254,254}{\strut m} \colorbox[RGB]{255,250,250}{\strut ood} \colorbox[RGB]{255,244,244}{\strut ,} \colorbox[RGB]{255,240,240}{\strut and} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,244,244}{\strut movie} \colorbox[RGB]{255,242,242}{\strut itself} \colorbox[RGB]{255,248,248}{\strut is} \colorbox[RGB]{255,238,238}{\strut really} \colorbox[RGB]{255,244,244}{\strut a} \colorbox[RGB]{255,208,208}{\strut great} \colorbox[RGB]{255,230,230}{\strut hit} \colorbox[RGB]{255,236,236}{\strut to} \colorbox[RGB]{255,224,224}{\strut me} \colorbox[RGB]{255,238,238}{\strut ,} \colorbox[RGB]{255,244,244}{\strut AL} \colorbox[RGB]{255,242,242}{\strut WA} \colorbox[RGB]{255,220,220}{\strut YS} \colorbox[RGB]{255,174,174}{\strut !} \colorbox[RGB]{255,238,238}{\strut R} \colorbox[RGB]{255,188,188}{\strut ating} \colorbox[RGB]{255,226,226}{\strut } \colorbox[RGB]{255,202,202}{\strut 4} \colorbox[RGB]{255,238,238}{\strut out} \colorbox[RGB]{255,226,226}{\strut of} \colorbox[RGB]{255,220,220}{\strut } \colorbox[RGB]{255,138,138}{\strut 5} \colorbox[RGB]{255,158,158}{\strut stars} \colorbox[RGB]{255,0,0}{\strut .}}}
\fbox{
\parbox{\textwidth}{
\setlength\fboxsep{0pt}
\colorbox[RGB]{255,0,0}{\strut <s>} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{252,252,255}{\strut went} \colorbox[RGB]{254,254,255}{\strut to} \colorbox[RGB]{252,252,255}{\strut see} \colorbox[RGB]{255,246,246}{\strut Ham} \colorbox[RGB]{255,224,224}{\strut let} \colorbox[RGB]{248,248,255}{\strut because} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,252,252}{\strut was} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{255,252,252}{\strut between} \colorbox[RGB]{243,243,255}{\strut jobs} \colorbox[RGB]{255,254,254}{\strut .} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{246,246,255}{\strut figured} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{248,248,255}{\strut 4} \colorbox[RGB]{252,252,255}{\strut hours} \colorbox[RGB]{255,250,250}{\strut would} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,238,238}{\strut great} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,252,252}{\strut '} \colorbox[RGB]{255,254,254}{\strut ve} \colorbox[RGB]{255,248,248}{\strut been} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{255,242,242}{\strut fan} \colorbox[RGB]{255,248,248}{\strut of} \colorbox[RGB]{230,230,255}{\strut Bran} \colorbox[RGB]{255,226,226}{\strut agh} \colorbox[RGB]{255,232,232}{\strut ;} \colorbox[RGB]{255,220,220}{\strut Dead} \colorbox[RGB]{255,252,252}{\strut Again} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,228,228}{\strut Henry} \colorbox[RGB]{255,254,254}{\strut V} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,244,244}{\strut I} \colorbox[RGB]{255,248,248}{\strut was} \colorbox[RGB]{255,254,254}{\strut completely} \colorbox[RGB]{246,246,255}{\strut over} \colorbox[RGB]{255,250,250}{\strut wh} \colorbox[RGB]{255,254,254}{\strut el} \colorbox[RGB]{254,254,255}{\strut med} \colorbox[RGB]{255,244,244}{\strut by} \colorbox[RGB]{255,248,248}{\strut the} \colorbox[RGB]{248,248,255}{\strut direction} \colorbox[RGB]{255,242,242}{\strut ,} \colorbox[RGB]{255,252,252}{\strut acting} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,244,244}{\strut cinemat} \colorbox[RGB]{255,254,254}{\strut ography} \colorbox[RGB]{255,246,246}{\strut that} \colorbox[RGB]{255,242,242}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,224,224}{\strut captured} \colorbox[RGB]{255,220,220}{\strut .} \colorbox[RGB]{216,216,255}{\strut Like} \colorbox[RGB]{255,250,250}{\strut other} \colorbox[RGB]{232,232,255}{\strut reviews} \colorbox[RGB]{255,246,246}{\strut the} \colorbox[RGB]{254,254,255}{\strut } \colorbox[RGB]{252,252,255}{\strut 4} \colorbox[RGB]{248,248,255}{\strut hours} \colorbox[RGB]{255,252,252}{\strut passes} \colorbox[RGB]{255,236,236}{\strut swift} \colorbox[RGB]{255,246,246}{\strut ly} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{255,195,195}{\strut Bran} \colorbox[RGB]{255,224,224}{\strut agh} \colorbox[RGB]{252,252,255}{\strut doesn} \colorbox[RGB]{255,242,242}{\strut '} \colorbox[RGB]{250,250,255}{\strut t} \colorbox[RGB]{252,252,255}{\strut play} \colorbox[RGB]{240,240,255}{\strut Ham} \colorbox[RGB]{255,252,252}{\strut let} \colorbox[RGB]{255,254,254}{\strut ,} \colorbox[RGB]{255,248,248}{\strut he} \colorbox[RGB]{255,250,250}{\strut is} \colorbox[RGB]{255,254,254}{\strut Ham} \colorbox[RGB]{255,252,252}{\strut let} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,240,240}{\strut born} \colorbox[RGB]{255,236,236}{\strut for} \colorbox[RGB]{255,240,240}{\strut this} \colorbox[RGB]{255,220,220}{\strut .} \colorbox[RGB]{255,244,244}{\strut When} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{255,252,252}{\strut watch} \colorbox[RGB]{255,240,240}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{254,254,255}{\strut I} \colorbox[RGB]{234,234,255}{\strut '} \colorbox[RGB]{252,252,255}{\strut m} \colorbox[RGB]{254,254,255}{\strut constantly} \colorbox[RGB]{254,254,255}{\strut trying} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,252,252}{\strut find} \colorbox[RGB]{255,250,250}{\strut fault} \colorbox[RGB]{255,252,252}{\strut s} \colorbox[RGB]{255,254,254}{\strut ,} \colorbox[RGB]{254,254,255}{\strut I} \colorbox[RGB]{248,248,255}{\strut '} \colorbox[RGB]{254,254,255}{\strut ve} \colorbox[RGB]{254,254,255}{\strut looked} \colorbox[RGB]{254,254,255}{\strut at} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut go} \colorbox[RGB]{252,252,255}{\strut of} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{252,252,255}{\strut haven} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{252,252,255}{\strut t} \colorbox[RGB]{252,252,255}{\strut noticed} \colorbox[RGB]{255,254,254}{\strut them} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{227,227,255}{\strut How} \colorbox[RGB]{255,246,246}{\strut he} \colorbox[RGB]{254,254,255}{\strut was} \colorbox[RGB]{255,252,252}{\strut able} \colorbox[RGB]{254,254,255}{\strut to} \colorbox[RGB]{252,252,255}{\strut move} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut camera} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,250,250}{\strut Hall} \colorbox[RGB]{255,250,250}{\strut with} \colorbox[RGB]{255,254,254}{\strut all} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,248,248}{\strut mirror} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{248,248,255}{\strut is} \colorbox[RGB]{252,252,255}{\strut a} \colorbox[RGB]{238,238,255}{\strut mystery} \colorbox[RGB]{252,252,255}{\strut to} \colorbox[RGB]{248,248,255}{\strut me} \colorbox[RGB]{246,246,255}{\strut .} \colorbox[RGB]{246,246,255}{\strut This} \colorbox[RGB]{240,240,255}{\strut movie} \colorbox[RGB]{250,250,255}{\strut was} \colorbox[RGB]{255,246,246}{\strut shot} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{255,252,252}{\strut 7} \colorbox[RGB]{255,254,254}{\strut 0} \colorbox[RGB]{255,248,248}{\strut mil} \colorbox[RGB]{255,244,244}{\strut .} \colorbox[RGB]{255,254,254}{\strut It} \colorbox[RGB]{255,230,230}{\strut '} \colorbox[RGB]{252,252,255}{\strut s} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{232,232,255}{\strut shame} \colorbox[RGB]{255,252,252}{\strut that} \colorbox[RGB]{255,238,238}{\strut Columbia} \colorbox[RGB]{248,248,255}{\strut hasn} \colorbox[RGB]{232,232,255}{\strut '} \colorbox[RGB]{252,252,255}{\strut t} \colorbox[RGB]{252,252,255}{\strut released} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut W} \colorbox[RGB]{255,250,250}{\strut ides} \colorbox[RGB]{255,240,240}{\strut creen} \colorbox[RGB]{255,252,252}{\strut version} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,246,246}{\strut this} \colorbox[RGB]{255,248,248}{\strut on} \colorbox[RGB]{255,238,238}{\strut V} \colorbox[RGB]{255,230,230}{\strut HS} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,254,254}{\strut own} \colorbox[RGB]{254,254,255}{\strut a} \colorbox[RGB]{255,254,254}{\strut DVD} \colorbox[RGB]{252,252,255}{\strut player} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut and} \colorbox[RGB]{255,244,244}{\strut I} \colorbox[RGB]{255,210,210}{\strut '} \colorbox[RGB]{254,254,255}{\strut d} \colorbox[RGB]{254,254,255}{\strut take} \colorbox[RGB]{255,246,246}{\strut this} \colorbox[RGB]{252,252,255}{\strut over} \colorbox[RGB]{254,254,255}{\strut T} \colorbox[RGB]{254,254,255}{\strut itan} \colorbox[RGB]{254,254,255}{\strut ic} \colorbox[RGB]{255,254,254}{\strut any} \colorbox[RGB]{255,254,254}{\strut day} \colorbox[RGB]{255,250,250}{\strut .} \colorbox[RGB]{255,250,250}{\strut So} \colorbox[RGB]{255,236,236}{\strut Columbia} \colorbox[RGB]{255,244,244}{\strut if} \colorbox[RGB]{255,244,244}{\strut you} \colorbox[RGB]{255,240,240}{\strut '} \colorbox[RGB]{254,254,255}{\strut re} \colorbox[RGB]{255,246,246}{\strut listening} \colorbox[RGB]{238,238,255}{\strut put} \colorbox[RGB]{255,246,246}{\strut this} \colorbox[RGB]{255,244,244}{\strut film} \colorbox[RGB]{252,252,255}{\strut out} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{252,252,255}{\strut way} \colorbox[RGB]{255,254,254}{\strut it} \colorbox[RGB]{252,252,255}{\strut should} \colorbox[RGB]{254,254,255}{\strut be} \colorbox[RGB]{255,252,252}{\strut watched} \colorbox[RGB]{255,97,97}{\strut !} \colorbox[RGB]{211,211,255}{\strut And} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{255,250,250}{\strut don} \colorbox[RGB]{255,234,234}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,244,244}{\strut know} \colorbox[RGB]{254,254,255}{\strut what} \colorbox[RGB]{254,254,255}{\strut happened} \colorbox[RGB]{255,254,254}{\strut at} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{252,252,255}{\strut O} \colorbox[RGB]{252,252,255}{\strut sc} \colorbox[RGB]{252,252,255}{\strut ars} \colorbox[RGB]{254,254,255}{\strut .} \colorbox[RGB]{255,254,254}{\strut This} \colorbox[RGB]{220,220,255}{\strut should} \colorbox[RGB]{234,234,255}{\strut have} \colorbox[RGB]{255,248,248}{\strut swe} \colorbox[RGB]{255,252,252}{\strut pt} \colorbox[RGB]{255,252,252}{\strut Best} \colorbox[RGB]{255,252,252}{\strut Picture} \colorbox[RGB]{255,246,246}{\strut ,} \colorbox[RGB]{254,254,255}{\strut Best} \colorbox[RGB]{255,254,254}{\strut A} \colorbox[RGB]{255,254,254}{\strut ctor} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Best} \colorbox[RGB]{255,254,254}{\strut D} \colorbox[RGB]{254,254,255}{\strut irection} \colorbox[RGB]{255,252,252}{\strut ,} \colorbox[RGB]{254,254,255}{\strut best} \colorbox[RGB]{255,252,252}{\strut cinemat} \colorbox[RGB]{255,254,254}{\strut ography} \colorbox[RGB]{238,238,255}{\strut .} \colorbox[RGB]{224,224,255}{\strut What} \colorbox[RGB]{255,254,254}{\strut films} \colorbox[RGB]{255,250,250}{\strut were} \colorbox[RGB]{254,254,255}{\strut they} \colorbox[RGB]{250,250,255}{\strut watching} \colorbox[RGB]{255,242,242}{\strut ?} \colorbox[RGB]{255,238,238}{\strut I} \colorbox[RGB]{255,234,234}{\strut felt} \colorbox[RGB]{238,238,255}{\strut sorry} \colorbox[RGB]{250,250,255}{\strut for} \colorbox[RGB]{255,226,226}{\strut Bran} \colorbox[RGB]{255,248,248}{\strut agh} \colorbox[RGB]{255,244,244}{\strut at} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,252,252}{\strut O} \colorbox[RGB]{255,250,250}{\strut sc} \colorbox[RGB]{255,246,246}{\strut ars} \colorbox[RGB]{255,216,216}{\strut when} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{255,252,252}{\strut did} \colorbox[RGB]{255,250,250}{\strut a} \colorbox[RGB]{254,254,255}{\strut t} \colorbox[RGB]{254,254,255}{\strut ribute} \colorbox[RGB]{255,246,246}{\strut to} \colorbox[RGB]{255,142,142}{\strut Shakespeare} \colorbox[RGB]{255,250,250}{\strut on} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{254,254,255}{\strut screen} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{226,226,255}{\strut They} \colorbox[RGB]{124,124,255}{\strut should} \colorbox[RGB]{176,176,255}{\strut have} \colorbox[RGB]{198,198,255}{\strut been} \colorbox[RGB]{255,252,252}{\strut giving} \colorbox[RGB]{255,250,250}{\strut a} \colorbox[RGB]{255,252,252}{\strut t} \colorbox[RGB]{255,234,234}{\strut ribute} \colorbox[RGB]{255,244,244}{\strut to} \colorbox[RGB]{255,208,208}{\strut Bran} \colorbox[RGB]{255,252,252}{\strut agh} \colorbox[RGB]{255,210,210}{\strut for} \colorbox[RGB]{255,179,179}{\strut bringing} \colorbox[RGB]{255,206,206}{\strut us} \colorbox[RGB]{255,238,238}{\strut one} \colorbox[RGB]{255,242,242}{\strut of} \colorbox[RGB]{255,248,248}{\strut the} \colorbox[RGB]{255,244,244}{\strut greatest} \colorbox[RGB]{255,244,244}{\strut films} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{254,254,255}{\strut all} \colorbox[RGB]{255,244,244}{\strut time} \colorbox[RGB]{255,44,44}{\strut .}}}
\fbox{
\parbox{\textwidth}{
\setlength\fboxsep{0pt}
\colorbox[RGB]{255,17,17}{\strut <s>} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{254,254,255}{\strut went} \colorbox[RGB]{255,252,252}{\strut to} \colorbox[RGB]{255,252,252}{\strut see} \colorbox[RGB]{255,224,224}{\strut Ham} \colorbox[RGB]{255,222,222}{\strut let} \colorbox[RGB]{255,252,252}{\strut because} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,252,252}{\strut was} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,252,252}{\strut between} \colorbox[RGB]{252,252,255}{\strut jobs} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{254,254,255}{\strut figured} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{254,254,255}{\strut 4} \colorbox[RGB]{255,254,254}{\strut hours} \colorbox[RGB]{255,250,250}{\strut would} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,240,240}{\strut great} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{255,248,248}{\strut '} \colorbox[RGB]{255,252,252}{\strut ve} \colorbox[RGB]{255,250,250}{\strut been} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{255,242,242}{\strut fan} \colorbox[RGB]{255,246,246}{\strut of} \colorbox[RGB]{255,244,244}{\strut Bran} \colorbox[RGB]{255,228,228}{\strut agh} \colorbox[RGB]{255,240,240}{\strut ;} \colorbox[RGB]{255,234,234}{\strut Dead} \colorbox[RGB]{255,250,250}{\strut Again} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,236,236}{\strut Henry} \colorbox[RGB]{255,252,252}{\strut V} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,248,248}{\strut was} \colorbox[RGB]{255,248,248}{\strut completely} \colorbox[RGB]{255,254,254}{\strut over} \colorbox[RGB]{255,248,248}{\strut wh} \colorbox[RGB]{255,254,254}{\strut el} \colorbox[RGB]{255,252,252}{\strut med} \colorbox[RGB]{255,244,244}{\strut by} \colorbox[RGB]{255,248,248}{\strut the} \colorbox[RGB]{255,254,254}{\strut direction} \colorbox[RGB]{255,244,244}{\strut ,} \colorbox[RGB]{255,252,252}{\strut acting} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,248,248}{\strut cinemat} \colorbox[RGB]{255,254,254}{\strut ography} \colorbox[RGB]{255,248,248}{\strut that} \colorbox[RGB]{255,244,244}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,234,234}{\strut captured} \colorbox[RGB]{255,222,222}{\strut .} \colorbox[RGB]{240,240,255}{\strut Like} \colorbox[RGB]{255,246,246}{\strut other} \colorbox[RGB]{252,252,255}{\strut reviews} \colorbox[RGB]{255,240,240}{\strut the} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{255,252,252}{\strut 4} \colorbox[RGB]{255,252,252}{\strut hours} \colorbox[RGB]{255,250,250}{\strut passes} \colorbox[RGB]{255,242,242}{\strut swift} \colorbox[RGB]{255,246,246}{\strut ly} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,222,222}{\strut Bran} \colorbox[RGB]{255,232,232}{\strut agh} \colorbox[RGB]{254,254,255}{\strut doesn} \colorbox[RGB]{255,242,242}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,254,254}{\strut play} \colorbox[RGB]{255,254,254}{\strut Ham} \colorbox[RGB]{255,250,250}{\strut let} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{255,252,252}{\strut is} \colorbox[RGB]{255,252,252}{\strut Ham} \colorbox[RGB]{255,250,250}{\strut let} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,252,252}{\strut he} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,244,244}{\strut born} \colorbox[RGB]{255,244,244}{\strut for} \colorbox[RGB]{255,246,246}{\strut this} \colorbox[RGB]{255,228,228}{\strut .} \colorbox[RGB]{255,246,246}{\strut When} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{255,252,252}{\strut watch} \colorbox[RGB]{255,244,244}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{254,254,255}{\strut '} \colorbox[RGB]{254,254,255}{\strut m} \colorbox[RGB]{255,252,252}{\strut constantly} \colorbox[RGB]{255,254,254}{\strut trying} \colorbox[RGB]{255,252,252}{\strut to} \colorbox[RGB]{255,252,252}{\strut find} \colorbox[RGB]{255,248,248}{\strut fault} \colorbox[RGB]{255,252,252}{\strut s} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,252,252}{\strut '} \colorbox[RGB]{255,254,254}{\strut ve} \colorbox[RGB]{255,254,254}{\strut looked} \colorbox[RGB]{255,254,254}{\strut at} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut go} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,252,252}{\strut s} \colorbox[RGB]{255,250,250}{\strut and} \colorbox[RGB]{255,254,254}{\strut haven} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{254,254,255}{\strut t} \colorbox[RGB]{255,252,252}{\strut noticed} \colorbox[RGB]{255,254,254}{\strut them} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{248,248,255}{\strut How} \colorbox[RGB]{255,246,246}{\strut he} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,248,248}{\strut able} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{254,254,255}{\strut move} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut camera} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,250,250}{\strut Hall} \colorbox[RGB]{255,250,250}{\strut with} \colorbox[RGB]{255,254,254}{\strut all} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,252,252}{\strut mirror} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,252,252}{\strut is} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{252,252,255}{\strut mystery} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,250,250}{\strut me} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{252,252,255}{\strut This} \colorbox[RGB]{254,254,255}{\strut movie} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,248,248}{\strut shot} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{255,252,252}{\strut 7} \colorbox[RGB]{255,252,252}{\strut 0} \colorbox[RGB]{255,246,246}{\strut mil} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{255,250,250}{\strut It} \colorbox[RGB]{255,232,232}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,254,254}{\strut shame} \colorbox[RGB]{255,250,250}{\strut that} \colorbox[RGB]{255,242,242}{\strut Columbia} \colorbox[RGB]{254,254,255}{\strut hasn} \colorbox[RGB]{255,254,254}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,252,252}{\strut released} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut W} \colorbox[RGB]{255,250,250}{\strut ides} \colorbox[RGB]{255,240,240}{\strut creen} \colorbox[RGB]{255,250,250}{\strut version} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,248,248}{\strut this} \colorbox[RGB]{255,248,248}{\strut on} \colorbox[RGB]{255,242,242}{\strut V} \colorbox[RGB]{255,232,232}{\strut HS} \colorbox[RGB]{255,242,242}{\strut .} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{255,250,250}{\strut own} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{255,248,248}{\strut DVD} \colorbox[RGB]{255,250,250}{\strut player} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,246,246}{\strut and} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,218,218}{\strut '} \colorbox[RGB]{255,254,254}{\strut d} \colorbox[RGB]{255,254,254}{\strut take} \colorbox[RGB]{255,248,248}{\strut this} \colorbox[RGB]{255,250,250}{\strut over} \colorbox[RGB]{255,254,254}{\strut T} \colorbox[RGB]{255,254,254}{\strut itan} \colorbox[RGB]{255,254,254}{\strut ic} \colorbox[RGB]{255,252,252}{\strut any} \colorbox[RGB]{255,252,252}{\strut day} \colorbox[RGB]{255,242,242}{\strut .} \colorbox[RGB]{255,246,246}{\strut So} \colorbox[RGB]{255,238,238}{\strut Columbia} \colorbox[RGB]{255,242,242}{\strut if} \colorbox[RGB]{255,246,246}{\strut you} \colorbox[RGB]{255,240,240}{\strut '} \colorbox[RGB]{255,254,254}{\strut re} \colorbox[RGB]{255,244,244}{\strut listening} \colorbox[RGB]{254,254,255}{\strut put} \colorbox[RGB]{255,244,244}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,252,252}{\strut out} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,246,246}{\strut way} \colorbox[RGB]{255,252,252}{\strut it} \colorbox[RGB]{254,254,255}{\strut should} \colorbox[RGB]{255,250,250}{\strut be} \colorbox[RGB]{255,248,248}{\strut watched} \colorbox[RGB]{255,152,152}{\strut !} \colorbox[RGB]{250,250,255}{\strut And} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,250,250}{\strut don} \colorbox[RGB]{255,234,234}{\strut '} \colorbox[RGB]{255,250,250}{\strut t} \colorbox[RGB]{255,246,246}{\strut know} \colorbox[RGB]{255,252,252}{\strut what} \colorbox[RGB]{255,250,250}{\strut happened} \colorbox[RGB]{255,252,252}{\strut at} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,252,252}{\strut O} \colorbox[RGB]{255,254,254}{\strut sc} \colorbox[RGB]{255,248,248}{\strut ars} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{255,248,248}{\strut This} \colorbox[RGB]{246,246,255}{\strut should} \colorbox[RGB]{255,252,252}{\strut have} \colorbox[RGB]{255,248,248}{\strut swe} \colorbox[RGB]{255,246,246}{\strut pt} \colorbox[RGB]{255,250,250}{\strut Best} \colorbox[RGB]{255,250,250}{\strut Picture} \colorbox[RGB]{255,242,242}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Best} \colorbox[RGB]{255,254,254}{\strut A} \colorbox[RGB]{255,252,252}{\strut ctor} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Best} \colorbox[RGB]{255,254,254}{\strut D} \colorbox[RGB]{255,254,254}{\strut irection} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut best} \colorbox[RGB]{255,252,252}{\strut cinemat} \colorbox[RGB]{255,252,252}{\strut ography} \colorbox[RGB]{255,238,238}{\strut .} \colorbox[RGB]{248,248,255}{\strut What} \colorbox[RGB]{255,242,242}{\strut films} \colorbox[RGB]{255,244,244}{\strut were} \colorbox[RGB]{255,246,246}{\strut they} \colorbox[RGB]{255,244,244}{\strut watching} \colorbox[RGB]{255,230,230}{\strut ?} \colorbox[RGB]{255,232,232}{\strut I} \colorbox[RGB]{255,234,234}{\strut felt} \colorbox[RGB]{255,254,254}{\strut sorry} \colorbox[RGB]{255,242,242}{\strut for} \colorbox[RGB]{255,230,230}{\strut Bran} \colorbox[RGB]{255,240,240}{\strut agh} \colorbox[RGB]{255,240,240}{\strut at} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,246,246}{\strut O} \colorbox[RGB]{255,248,248}{\strut sc} \colorbox[RGB]{255,240,240}{\strut ars} \colorbox[RGB]{255,220,220}{\strut when} \colorbox[RGB]{255,244,244}{\strut he} \colorbox[RGB]{255,248,248}{\strut did} \colorbox[RGB]{255,248,248}{\strut a} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,244,244}{\strut ribute} \colorbox[RGB]{255,238,238}{\strut to} \colorbox[RGB]{255,184,184}{\strut Shakespeare} \colorbox[RGB]{255,244,244}{\strut on} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,248,248}{\strut screen} \colorbox[RGB]{255,222,222}{\strut .} \colorbox[RGB]{255,252,252}{\strut They} \colorbox[RGB]{220,220,255}{\strut should} \colorbox[RGB]{255,234,234}{\strut have} \colorbox[RGB]{255,230,230}{\strut been} \colorbox[RGB]{255,220,220}{\strut giving} \colorbox[RGB]{255,244,244}{\strut a} \colorbox[RGB]{255,250,250}{\strut t} \colorbox[RGB]{255,222,222}{\strut ribute} \colorbox[RGB]{255,226,226}{\strut to} \colorbox[RGB]{255,224,224}{\strut Bran} \colorbox[RGB]{255,234,234}{\strut agh} \colorbox[RGB]{255,204,204}{\strut for} \colorbox[RGB]{255,200,200}{\strut bringing} \colorbox[RGB]{255,204,204}{\strut us} \colorbox[RGB]{255,224,224}{\strut one} \colorbox[RGB]{255,230,230}{\strut of} \colorbox[RGB]{255,236,236}{\strut the} \colorbox[RGB]{255,226,226}{\strut greatest} \colorbox[RGB]{255,228,228}{\strut films} \colorbox[RGB]{255,242,242}{\strut of} \colorbox[RGB]{255,240,240}{\strut all} \colorbox[RGB]{255,214,214}{\strut time} \colorbox[RGB]{255,0,0}{\strut .}}}
\fbox{
\parbox{\textwidth}{
\setlength\fboxsep{0pt}
\colorbox[RGB]{255,17,17}{\strut <s>} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{254,254,255}{\strut went} \colorbox[RGB]{255,252,252}{\strut to} \colorbox[RGB]{255,252,252}{\strut see} \colorbox[RGB]{255,224,224}{\strut Ham} \colorbox[RGB]{255,222,222}{\strut let} \colorbox[RGB]{255,252,252}{\strut because} \colorbox[RGB]{255,252,252}{\strut I} \colorbox[RGB]{255,252,252}{\strut was} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,252,252}{\strut between} \colorbox[RGB]{252,252,255}{\strut jobs} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{254,254,255}{\strut figured} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{254,254,255}{\strut 4} \colorbox[RGB]{255,254,254}{\strut hours} \colorbox[RGB]{255,250,250}{\strut would} \colorbox[RGB]{255,254,254}{\strut be} \colorbox[RGB]{255,240,240}{\strut great} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,250,250}{\strut I} \colorbox[RGB]{255,248,248}{\strut '} \colorbox[RGB]{255,252,252}{\strut ve} \colorbox[RGB]{255,250,250}{\strut been} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{255,242,242}{\strut fan} \colorbox[RGB]{255,246,246}{\strut of} \colorbox[RGB]{255,244,244}{\strut Bran} \colorbox[RGB]{255,228,228}{\strut agh} \colorbox[RGB]{255,240,240}{\strut ;} \colorbox[RGB]{255,234,234}{\strut Dead} \colorbox[RGB]{255,250,250}{\strut Again} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,236,236}{\strut Henry} \colorbox[RGB]{255,252,252}{\strut V} \colorbox[RGB]{255,246,246}{\strut .} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,248,248}{\strut was} \colorbox[RGB]{255,248,248}{\strut completely} \colorbox[RGB]{255,254,254}{\strut over} \colorbox[RGB]{255,248,248}{\strut wh} \colorbox[RGB]{255,254,254}{\strut el} \colorbox[RGB]{255,252,252}{\strut med} \colorbox[RGB]{255,244,244}{\strut by} \colorbox[RGB]{255,248,248}{\strut the} \colorbox[RGB]{255,254,254}{\strut direction} \colorbox[RGB]{255,244,244}{\strut ,} \colorbox[RGB]{255,252,252}{\strut acting} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,248,248}{\strut cinemat} \colorbox[RGB]{255,254,254}{\strut ography} \colorbox[RGB]{255,248,248}{\strut that} \colorbox[RGB]{255,244,244}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,234,234}{\strut captured} \colorbox[RGB]{255,222,222}{\strut .} \colorbox[RGB]{240,240,255}{\strut Like} \colorbox[RGB]{255,246,246}{\strut other} \colorbox[RGB]{252,252,255}{\strut reviews} \colorbox[RGB]{255,240,240}{\strut the} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{255,252,252}{\strut 4} \colorbox[RGB]{255,252,252}{\strut hours} \colorbox[RGB]{255,250,250}{\strut passes} \colorbox[RGB]{255,242,242}{\strut swift} \colorbox[RGB]{255,246,246}{\strut ly} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{255,222,222}{\strut Bran} \colorbox[RGB]{255,232,232}{\strut agh} \colorbox[RGB]{254,254,255}{\strut doesn} \colorbox[RGB]{255,242,242}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,254,254}{\strut play} \colorbox[RGB]{255,254,254}{\strut Ham} \colorbox[RGB]{255,250,250}{\strut let} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut he} \colorbox[RGB]{255,252,252}{\strut is} \colorbox[RGB]{255,252,252}{\strut Ham} \colorbox[RGB]{255,250,250}{\strut let} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,252,252}{\strut he} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,244,244}{\strut born} \colorbox[RGB]{255,244,244}{\strut for} \colorbox[RGB]{255,246,246}{\strut this} \colorbox[RGB]{255,228,228}{\strut .} \colorbox[RGB]{255,246,246}{\strut When} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{255,252,252}{\strut watch} \colorbox[RGB]{255,244,244}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{254,254,255}{\strut '} \colorbox[RGB]{254,254,255}{\strut m} \colorbox[RGB]{255,252,252}{\strut constantly} \colorbox[RGB]{255,254,254}{\strut trying} \colorbox[RGB]{255,252,252}{\strut to} \colorbox[RGB]{255,252,252}{\strut find} \colorbox[RGB]{255,248,248}{\strut fault} \colorbox[RGB]{255,252,252}{\strut s} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,254,254}{\strut I} \colorbox[RGB]{255,252,252}{\strut '} \colorbox[RGB]{255,254,254}{\strut ve} \colorbox[RGB]{255,254,254}{\strut looked} \colorbox[RGB]{255,254,254}{\strut at} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,254,254}{\strut go} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,252,252}{\strut s} \colorbox[RGB]{255,250,250}{\strut and} \colorbox[RGB]{255,254,254}{\strut haven} \colorbox[RGB]{255,250,250}{\strut '} \colorbox[RGB]{254,254,255}{\strut t} \colorbox[RGB]{255,252,252}{\strut noticed} \colorbox[RGB]{255,254,254}{\strut them} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{248,248,255}{\strut How} \colorbox[RGB]{255,246,246}{\strut he} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,248,248}{\strut able} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{254,254,255}{\strut move} \colorbox[RGB]{255,254,254}{\strut the} \colorbox[RGB]{255,254,254}{\strut camera} \colorbox[RGB]{255,254,254}{\strut in} \colorbox[RGB]{255,252,252}{\strut and} \colorbox[RGB]{255,254,254}{\strut out} \colorbox[RGB]{255,254,254}{\strut of} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,250,250}{\strut Hall} \colorbox[RGB]{255,250,250}{\strut with} \colorbox[RGB]{255,254,254}{\strut all} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,252,252}{\strut mirror} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,252,252}{\strut is} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{252,252,255}{\strut mystery} \colorbox[RGB]{255,254,254}{\strut to} \colorbox[RGB]{255,250,250}{\strut me} \colorbox[RGB]{255,248,248}{\strut .} \colorbox[RGB]{252,252,255}{\strut This} \colorbox[RGB]{254,254,255}{\strut movie} \colorbox[RGB]{255,254,254}{\strut was} \colorbox[RGB]{255,248,248}{\strut shot} \colorbox[RGB]{255,252,252}{\strut in} \colorbox[RGB]{255,254,254}{\strut } \colorbox[RGB]{255,252,252}{\strut 7} \colorbox[RGB]{255,252,252}{\strut 0} \colorbox[RGB]{255,246,246}{\strut mil} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{255,250,250}{\strut It} \colorbox[RGB]{255,232,232}{\strut '} \colorbox[RGB]{255,254,254}{\strut s} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,254,254}{\strut shame} \colorbox[RGB]{255,250,250}{\strut that} \colorbox[RGB]{255,242,242}{\strut Columbia} \colorbox[RGB]{254,254,255}{\strut hasn} \colorbox[RGB]{255,254,254}{\strut '} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,252,252}{\strut released} \colorbox[RGB]{255,254,254}{\strut a} \colorbox[RGB]{255,252,252}{\strut W} \colorbox[RGB]{255,250,250}{\strut ides} \colorbox[RGB]{255,240,240}{\strut creen} \colorbox[RGB]{255,250,250}{\strut version} \colorbox[RGB]{255,252,252}{\strut of} \colorbox[RGB]{255,248,248}{\strut this} \colorbox[RGB]{255,248,248}{\strut on} \colorbox[RGB]{255,242,242}{\strut V} \colorbox[RGB]{255,232,232}{\strut HS} \colorbox[RGB]{255,242,242}{\strut .} \colorbox[RGB]{255,248,248}{\strut I} \colorbox[RGB]{255,250,250}{\strut own} \colorbox[RGB]{255,252,252}{\strut a} \colorbox[RGB]{255,248,248}{\strut DVD} \colorbox[RGB]{255,250,250}{\strut player} \colorbox[RGB]{255,248,248}{\strut ,} \colorbox[RGB]{255,246,246}{\strut and} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,218,218}{\strut '} \colorbox[RGB]{255,254,254}{\strut d} \colorbox[RGB]{255,254,254}{\strut take} \colorbox[RGB]{255,248,248}{\strut this} \colorbox[RGB]{255,250,250}{\strut over} \colorbox[RGB]{255,254,254}{\strut T} \colorbox[RGB]{255,254,254}{\strut itan} \colorbox[RGB]{255,254,254}{\strut ic} \colorbox[RGB]{255,252,252}{\strut any} \colorbox[RGB]{255,252,252}{\strut day} \colorbox[RGB]{255,242,242}{\strut .} \colorbox[RGB]{255,246,246}{\strut So} \colorbox[RGB]{255,238,238}{\strut Columbia} \colorbox[RGB]{255,242,242}{\strut if} \colorbox[RGB]{255,246,246}{\strut you} \colorbox[RGB]{255,240,240}{\strut '} \colorbox[RGB]{255,254,254}{\strut re} \colorbox[RGB]{255,244,244}{\strut listening} \colorbox[RGB]{254,254,255}{\strut put} \colorbox[RGB]{255,244,244}{\strut this} \colorbox[RGB]{255,246,246}{\strut film} \colorbox[RGB]{255,252,252}{\strut out} \colorbox[RGB]{255,252,252}{\strut the} \colorbox[RGB]{255,246,246}{\strut way} \colorbox[RGB]{255,252,252}{\strut it} \colorbox[RGB]{254,254,255}{\strut should} \colorbox[RGB]{255,250,250}{\strut be} \colorbox[RGB]{255,248,248}{\strut watched} \colorbox[RGB]{255,152,152}{\strut !} \colorbox[RGB]{250,250,255}{\strut And} \colorbox[RGB]{255,246,246}{\strut I} \colorbox[RGB]{255,250,250}{\strut don} \colorbox[RGB]{255,234,234}{\strut '} \colorbox[RGB]{255,250,250}{\strut t} \colorbox[RGB]{255,246,246}{\strut know} \colorbox[RGB]{255,252,252}{\strut what} \colorbox[RGB]{255,250,250}{\strut happened} \colorbox[RGB]{255,252,252}{\strut at} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,252,252}{\strut O} \colorbox[RGB]{255,254,254}{\strut sc} \colorbox[RGB]{255,248,248}{\strut ars} \colorbox[RGB]{255,240,240}{\strut .} \colorbox[RGB]{255,248,248}{\strut This} \colorbox[RGB]{246,246,255}{\strut should} \colorbox[RGB]{255,252,252}{\strut have} \colorbox[RGB]{255,248,248}{\strut swe} \colorbox[RGB]{255,246,246}{\strut pt} \colorbox[RGB]{255,250,250}{\strut Best} \colorbox[RGB]{255,250,250}{\strut Picture} \colorbox[RGB]{255,242,242}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Best} \colorbox[RGB]{255,254,254}{\strut A} \colorbox[RGB]{255,252,252}{\strut ctor} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,254,254}{\strut Best} \colorbox[RGB]{255,254,254}{\strut D} \colorbox[RGB]{255,254,254}{\strut irection} \colorbox[RGB]{255,250,250}{\strut ,} \colorbox[RGB]{255,250,250}{\strut best} \colorbox[RGB]{255,252,252}{\strut cinemat} \colorbox[RGB]{255,252,252}{\strut ography} \colorbox[RGB]{255,238,238}{\strut .} \colorbox[RGB]{248,248,255}{\strut What} \colorbox[RGB]{255,242,242}{\strut films} \colorbox[RGB]{255,244,244}{\strut were} \colorbox[RGB]{255,246,246}{\strut they} \colorbox[RGB]{255,244,244}{\strut watching} \colorbox[RGB]{255,230,230}{\strut ?} \colorbox[RGB]{255,232,232}{\strut I} \colorbox[RGB]{255,234,234}{\strut felt} \colorbox[RGB]{255,254,254}{\strut sorry} \colorbox[RGB]{255,242,242}{\strut for} \colorbox[RGB]{255,230,230}{\strut Bran} \colorbox[RGB]{255,240,240}{\strut agh} \colorbox[RGB]{255,240,240}{\strut at} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,246,246}{\strut O} \colorbox[RGB]{255,248,248}{\strut sc} \colorbox[RGB]{255,240,240}{\strut ars} \colorbox[RGB]{255,220,220}{\strut when} \colorbox[RGB]{255,244,244}{\strut he} \colorbox[RGB]{255,248,248}{\strut did} \colorbox[RGB]{255,248,248}{\strut a} \colorbox[RGB]{255,254,254}{\strut t} \colorbox[RGB]{255,244,244}{\strut ribute} \colorbox[RGB]{255,238,238}{\strut to} \colorbox[RGB]{255,184,184}{\strut Shakespeare} \colorbox[RGB]{255,244,244}{\strut on} \colorbox[RGB]{255,250,250}{\strut the} \colorbox[RGB]{255,248,248}{\strut screen} \colorbox[RGB]{255,222,222}{\strut .} \colorbox[RGB]{255,252,252}{\strut They} \colorbox[RGB]{220,220,255}{\strut should} \colorbox[RGB]{255,234,234}{\strut have} \colorbox[RGB]{255,230,230}{\strut been} \colorbox[RGB]{255,220,220}{\strut giving} \colorbox[RGB]{255,244,244}{\strut a} \colorbox[RGB]{255,250,250}{\strut t} \colorbox[RGB]{255,222,222}{\strut ribute} \colorbox[RGB]{255,226,226}{\strut to} \colorbox[RGB]{255,224,224}{\strut Bran} \colorbox[RGB]{255,234,234}{\strut agh} \colorbox[RGB]{255,204,204}{\strut for} \colorbox[RGB]{255,200,200}{\strut bringing} \colorbox[RGB]{255,204,204}{\strut us} \colorbox[RGB]{255,224,224}{\strut one} \colorbox[RGB]{255,230,230}{\strut of} \colorbox[RGB]{255,236,236}{\strut the} \colorbox[RGB]{255,226,226}{\strut greatest} \colorbox[RGB]{255,228,228}{\strut films} \colorbox[RGB]{255,242,242}{\strut of} \colorbox[RGB]{255,240,240}{\strut all} \colorbox[RGB]{255,214,214}{\strut time} \colorbox[RGB]{255,0,0}{\strut .}}}
\begin{thebibliography}{52}
\providecommand{\natexlab}[1]{#1}
\providecommand{\url}[1]{\texttt{#1}}
\expandafter\ifx\csname urlstyle\endcsname\relax
\providecommand{\doi}[1]{doi: #1}\else
\providecommand{\doi}{doi: \begingroup \urlstyle{rm}\Url}\fi
\bibitem[Abnar and Zuidema(2020)]{abnar2020quantifying}
Samira Abnar and Willem Zuidema.
\newblock Quantifying attention flow in transformers.
\newblock In \emph{Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics}. Association for Computational Linguistics, 2020.
\bibitem[Achtibat et~al.(2024)Achtibat, Hatefi, Dreyer, Jain, Wiegand, Lapuschkin, and Samek]{achtibat2024attnlrp}
Reduan Achtibat, Sayed Mohammad~Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek.
\newblock Attnlrp: attention-aware layer-wise relevance propagation for transformers.
\newblock In \emph{Proceedings of the 41st International Conference on Machine Learning}, pages 135--168, 2024.
\bibitem[Ali et~al.(2022{\natexlab{a}})Ali, Schnake, Eberle, Montavon, M{\"u}ller, and Wolf]{ali2022xai}
Ameen Ali, Thomas Schnake, Oliver Eberle, Gr{\'e}goire Montavon, Klaus-Robert M{\"u}ller, and Lior Wolf.
\newblock Xai for transformers: Better explanations through conservative propagation.
\newblock In \emph{International conference on machine learning}, pages 435--451. PMLR, 2022{\natexlab{a}}.
\bibitem[Ali et~al.(2022{\natexlab{b}})Ali, Schwartz, Hazan, and Wolf]{ali2022video}
Ameen Ali, Idan Schwartz, Tamir Hazan, and Lior Wolf.
\newblock Video and text matching with conditioned embeddings.
\newblock In \emph{Proceedings of the IEEE/CVF winter conference on applications of computer vision}, pages 1565--1574, 2022{\natexlab{b}}.
\bibitem[Ali et~al.(2024)Ali, Zimerman, and Wolf]{ali2024hidden}
Ameen Ali, Itamar Zimerman, and Lior Wolf.
\newblock The hidden attention of mamba models.
\newblock \emph{arXiv preprint arXiv:2403.01590}, 2024.
\bibitem[Arras et~al.(2017)Arras, Montavon, M{\"u}ller, and Samek]{arras-etal-2017-explaining}
Leila Arras, Gr{\'e}goire Montavon, Klaus-Robert M{\"u}ller, and Wojciech Samek.
\newblock Explaining recurrent neural network predictions in sentiment analysis.
\newblock In \emph{Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis}, pages 159--168, 2017.
\bibitem[Arras et~al.(2019)Arras, Arjona-Medina, Widrich, Montavon, Gillhofer, M{\"u}ller, Hochreiter, and Samek]{arras2019explaining}
Leila Arras, Jos{\'e} Arjona-Medina, Michael Widrich, Gr{\'e}goire Montavon, Michael Gillhofer, Klaus-Robert M{\"u}ller, Sepp Hochreiter, and Wojciech Samek.
\newblock Explaining and interpreting lstms.
\newblock \emph{Explainable ai: Interpreting, explaining and visualizing deep learning}, pages 211--238, 2019.
\bibitem[Arrieta et~al.(2020)Arrieta, D{\'\i}az-Rodr{\'\i}guez, Del~Ser, Bennetot, Tabik, Barbado, Garc{\'\i}a, Gil-L{\'o}pez, Molina, Benjamins, et~al.]{arrieta2020explainable}
Alejandro~Barredo Arrieta, Natalia D{\'\i}az-Rodr{\'\i}guez, Javier Del~Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garc{\'\i}a, Sergio Gil-L{\'o}pez, Daniel Molina, Richard Benjamins, et~al.
\newblock Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai.
\newblock \emph{Information fusion}, 58:\penalty0 82--115, 2020.
\bibitem[Bach et~al.(2015)Bach, Binder, Montavon, Klauschen, M{\"u}ller, and Samek]{bach2015pixel}
Sebastian Bach, Alexander Binder, Gr{\'e}goire Montavon, Frederick Klauschen, Klaus-Robert M{\"u}ller, and Wojciech Samek.
\newblock On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.
\newblock \emph{PloS one}, 10\penalty0 (7):\penalty0 e0130140, 2015.
\bibitem[Baehrens et~al.(2010)Baehrens, Schroeter, Harmeling, Kawanabe, Hansen, and M{\"u}ller]{baehrens2010explain}
D~Baehrens, T~Schroeter, S~Harmeling, M~Kawanabe, K~Hansen, and K-R M{\"u}ller.
\newblock How to explain individual classification decisions.
\newblock \emph{Journal of Machine Learning Research}, 2010.
\bibitem[Bai et~al.(2023)Bai, Bai, Chu, Cui, Dang, Deng, Fan, Ge, Han, Huang, et~al.]{bai2023qwen}
Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu~Han, Fei Huang, et~al.
\newblock Qwen technical report.
\newblock \emph{arXiv preprint arXiv:2309.16609}, 2023.
\bibitem[Balduzzi et~al.(2017)Balduzzi, Frean, Leary, Lewis, Ma, and McWilliams]{balduzzi2017shattered}
David Balduzzi, Marcus Frean, Lennox Leary, JP~Lewis, Kurt Wan-Duo Ma, and Brian McWilliams.
\newblock The shattered gradients problem: If resnets are the answer, then what is the question?
\newblock In \emph{International conference on machine learning}, pages 342--350. PMLR, 2017.
\bibitem[Biderman et~al.(2023)Biderman, Schoelkopf, Anthony, Bradley, O’Brien, Hallahan, Khan, Purohit, Prashanth, Raff, et~al.]{biderman2023pythia}
Stella Biderman, Hailey Schoelkopf, Quentin~Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad~Aflah Khan, Shivanshu Purohit, USVSN~Sai Prashanth, Edward Raff, et~al.
\newblock Pythia: A suite for analyzing large language models across training and scaling.
\newblock In \emph{International Conference on Machine Learning}, pages 2397--2430. PMLR, 2023.
\bibitem[Caron et~al.(2021)Caron, Touvron, Misra, J{\'e}gou, Mairal, Bojanowski, and Joulin]{caron2021emerging}
Mathilde Caron, Hugo Touvron, Ishan Misra, Herv{\'e} J{\'e}gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin.
\newblock Emerging properties in self-supervised vision transformers.
\newblock In \emph{Proceedings of the IEEE/CVF international conference on computer vision}, pages 9650--9660, 2021.
\bibitem[Chefer et~al.(2021{\natexlab{a}})Chefer, Gur, and Wolf]{chefer2021generic}
Hila Chefer, Shir Gur, and Lior Wolf.
\newblock Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers.
\newblock In \emph{Proceedings of the IEEE/CVF international conference on computer vision}, pages 397--406, 2021{\natexlab{a}}.
\bibitem[Chefer et~al.(2021{\natexlab{b}})Chefer, Gur, and Wolf]{chefer2021transformer}
Hila Chefer, Shir Gur, and Lior Wolf.
\newblock Transformer interpretability beyond attention visualization.
\newblock In \emph{Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, pages 782--791, 2021{\natexlab{b}}.
\bibitem[Cheng et~al.(2025)Cheng, Xing, Xue, and Yang]{cheng2025unifying}
Qisen Cheng, Jinming Xing, Chang Xue, and Xiaoran Yang.
\newblock Unifying prediction and explanation in time-series transformers via shapley-based pretraining.
\newblock \emph{arXiv preprint arXiv:2501.15070}, 2025.
\bibitem[Clark et~al.(2019)Clark, Khandelwal, Levy, and Manning]{clark2019does}
Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher~D Manning.
\newblock What does bert look at? an analysis of bert’s attention.
\newblock In \emph{Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP}, page 276. Association for Computational Linguistics, 2019.
\bibitem[Das and Rad(2020)]{das2020opportunities}
Arun Das and Paul Rad.
\newblock Opportunities and challenges in explainable artificial intelligence (xai): A survey.
\newblock \emph{arXiv preprint arXiv:2006.11371}, 2020.
\bibitem[Dufter et~al.(2022)Dufter, Schmitt, and Sch{\"u}tze]{dufter2022position}
Philipp Dufter, Martin Schmitt, and Hinrich Sch{\"u}tze.
\newblock Position information in transformers: An overview.
\newblock \emph{Computational Linguistics}, 48\penalty0 (3):\penalty0 733--763, 2022.
\bibitem[Grattafiori et~al.(2024)Grattafiori, Dubey, Jauhri, Pandey, Kadian, Al-Dahle, Letman, Mathur, Schelten, Vaughan, et~al.]{grattafiori2024llama}
Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et~al.
\newblock The llama 3 herd of models.
\newblock \emph{arXiv preprint arXiv:2407.21783}, 2024.
\bibitem[Guillaumin et~al.(2014)Guillaumin, K{\"u}ttel, and Ferrari]{guillaumin2014imagenet}
Matthieu Guillaumin, Daniel K{\"u}ttel, and Vittorio Ferrari.
\newblock Imagenet auto-annotation with segmentation propagation.
\newblock \emph{International Journal of Computer Vision}, 110:\penalty0 328 -- 348, 2014.
\newblock URL \url{https://api.semanticscholar.org/CorpusID:1005559}.
\bibitem[Hochreiter and Schmidhuber(1997)]{hochreiter1997long}
Sepp Hochreiter and J{\"u}rgen Schmidhuber.
\newblock Long short-term memory.
\newblock \emph{Neural computation}, 9\penalty0 (8):\penalty0 1735--1780, 1997.
\bibitem[Ibrahim and Shafiq(2023)]{ibrahim2023explainable}
Rami Ibrahim and M~Omair Shafiq.
\newblock Explainable convolutional neural networks: a taxonomy, review, and future directions.
\newblock \emph{ACM Computing Surveys}, 55\penalty0 (10):\penalty0 1--37, 2023.
\bibitem[Jafari et~al.(2024)Jafari, Montavon, Müller, and Eberle]{jafari2024mambalrp}
Farnoush~Rezaei Jafari, Grégoire Montavon, Klaus-Robert Müller, and Oliver Eberle.
\newblock Mambalrp: Explaining selective state space sequence models.
\newblock In \emph{The Thirty-eighth Annual Conference on Neural Information Processing Systems}, 2024.
\bibitem[Jain and Wallace(2019)]{jain2019attention}
Sarthak Jain and Byron~C Wallace.
\newblock Attention is not explanation.
\newblock In \emph{Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)}, pages 3543--3556, 2019.
\bibitem[Jordan(1997)]{jordan1997serial}
Michael~I Jordan.
\newblock Serial order: A parallel distributed processing approach.
\newblock In \emph{Advances in psychology}, volume 121, pages 471--495. Elsevier, 1997.
\bibitem[Kokalj et~al.(2021)Kokalj, {\v{S}}krlj, Lavra{\v{c}}, Pollak, and Robnik-{\v{S}}ikonja]{kokalj2021bert}
Enja Kokalj, Bla{\v{z}} {\v{S}}krlj, Nada Lavra{\v{c}}, Senja Pollak, and Marko Robnik-{\v{S}}ikonja.
\newblock Bert meets shapley: Extending shap explanations to transformer-based classifiers.
\newblock In \emph{Proceedings of the EACL hackashop on news media content analysis and automated report generation}, pages 16--21, 2021.
\bibitem[Lang et~al.(2021)Lang, Gandelsman, Yarom, Wald, Elidan, Hassidim, Freeman, Isola, Globerson, Irani, et~al.]{lang2021explaining}
Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William~T Freeman, Phillip Isola, Amir Globerson, Michal Irani, et~al.
\newblock Explaining in style: training a gan to explain a classifier in stylespace.
\newblock In \emph{Proceedings of the IEEE/CVF International Conference on Computer Vision}, pages 693--702, 2021.
\bibitem[Linardatos et~al.(2020)Linardatos, Papastefanopoulos, and Kotsiantis]{linardatos2020explainable}
Pantelis Linardatos, Vasilis Papastefanopoulos, and Sotiris Kotsiantis.
\newblock Explainable ai: A review of machine learning interpretability methods.
\newblock \emph{Entropy}, 23\penalty0 (1):\penalty0 18, 2020.
\bibitem[Montavon et~al.(2017)Montavon, Lapuschkin, Binder, Samek, and M{\"u}ller]{montavon2017explaining}
Gr{\'e}goire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert M{\"u}ller.
\newblock Explaining nonlinear classification decisions with deep taylor decomposition.
\newblock \emph{Pattern recognition}, 65:\penalty0 211--222, 2017.
\bibitem[Montavon et~al.(2019)Montavon, Binder, Lapuschkin, Samek, and M{\"u}ller]{montavon2019layer}
Gr{\'e}goire Montavon, Alexander Binder, Sebastian Lapuschkin, Wojciech Samek, and Klaus-Robert M{\"u}ller.
\newblock Layer-wise relevance propagation: an overview.
\newblock \emph{Explainable AI: interpreting, explaining and visualizing deep learning}, pages 193--209, 2019.
\bibitem[Mosca et~al.(2022)Mosca, Szigeti, Tragianni, Gallagher, and Groh]{mosca2022shap}
Edoardo Mosca, Ferenc Szigeti, Stella Tragianni, Daniel Gallagher, and Georg Groh.
\newblock Shap-based explanation methods: a review for nlp interpretability.
\newblock In \emph{Proceedings of the 29th international conference on computational linguistics}, pages 4593--4603, 2022.
\bibitem[Press et~al.()Press, Smith, and Lewis]{presstrain}
Ofir Press, Noah Smith, and Mike Lewis.
\newblock Train short, test long: Attention with linear biases enables input length extrapolation.
\newblock In \emph{International Conference on Learning Representations}.
\bibitem[Raffel et~al.(2020)Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li, and Liu]{raffel2020exploring}
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter~J Liu.
\newblock Exploring the limits of transfer learning with a unified text-to-text transformer.
\newblock \emph{Journal of machine learning research}, 21\penalty0 (140):\penalty0 1--67, 2020.
\bibitem[Ravi et~al.(2024)Ravi, Gabeur, Hu, Hu, Ryali, Ma, Khedr, R{\"a}dle, Rolland, Gustafson, et~al.]{ravi2024sam}
Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R{\"a}dle, Chloe Rolland, Laura Gustafson, et~al.
\newblock Sam 2: Segment anything in images and videos.
\newblock \emph{arXiv preprint arXiv:2408.00714}, 2024.
\bibitem[Samek et~al.(2017)Samek, Wiegand, and M{\"u}ller]{samek2017explainable}
Wojciech Samek, Thomas Wiegand, and Klaus-Robert M{\"u}ller.
\newblock Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models.
\newblock \emph{arXiv preprint arXiv:1708.08296}, 2017.
\bibitem[Selvaraju et~al.(2017)Selvaraju, Cogswell, Das, Vedantam, Parikh, and Batra]{selvaraju2017grad}
Ramprasaath~R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra.
\newblock Grad-cam: Visual explanations from deep networks via gradient-based localization.
\newblock In \emph{Proceedings of the IEEE international conference on computer vision}, pages 618--626, 2017.
\bibitem[Shaw et~al.(2018)Shaw, Uszkoreit, and Vaswani]{shaw2018self}
Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani.
\newblock Self-attention with relative position representations.
\newblock \emph{arXiv preprint arXiv:1803.02155}, 2018.
\bibitem[Shrikumar et~al.(2017)Shrikumar, Greenside, and Kundaje]{shrikumar2017learning}
Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje.
\newblock Learning important features through propagating activation differences.
\newblock In \emph{International conference on machine learning}, pages 3145--3153. PMLR, 2017.
\bibitem[Su et~al.(2024)Su, Ahmed, Lu, Pan, Bo, and Liu]{su2024roformer}
Jianlin Su, Murtadha Ahmed, Yu~Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu.
\newblock Roformer: Enhanced transformer with rotary position embedding.
\newblock \emph{Neurocomputing}, 568:\penalty0 127063, 2024.
\bibitem[Team et~al.(2024)Team, Mesnard, Hardin, Dadashi, Bhupatiraju, Pathak, Sifre, Rivi{\`e}re, Kale, Love, et~al.]{team2024gemma}
Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi{\`e}re, Mihir~Sanjay Kale, Juliette Love, et~al.
\newblock Gemma: Open models based on gemini research and technology.
\newblock \emph{arXiv preprint arXiv:2403.08295}, 2024.
\bibitem[Touvron et~al.(2021)Touvron, Cord, Douze, Massa, Sablayrolles, and J{\'e}gou]{touvron2021training}
Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv{\'e} J{\'e}gou.
\newblock Training data-efficient image transformers \& distillation through attention.
\newblock In \emph{International conference on machine learning}, pages 10347--10357. PMLR, 2021.
\bibitem[Touvron et~al.(2023)Touvron, Lavril, Izacard, Martinet, Lachaux, Lacroix, Rozi{\`e}re, Goyal, Hambro, Azhar, et~al.]{touvron2023llama}
Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth{\'e}e Lacroix, Baptiste Rozi{\`e}re, Naman Goyal, Eric Hambro, Faisal Azhar, et~al.
\newblock Llama: Open and efficient foundation language models.
\newblock \emph{arXiv preprint arXiv:2302.13971}, 2023.
\bibitem[Vaswani(2017)]{vaswani2017attention}
A~Vaswani.
\newblock Attention is all you need.
\newblock \emph{Advances in Neural Information Processing Systems}, 2017.
\bibitem[Voita et~al.(2021)Voita, Sennrich, and Titov]{voita2021analyzing}
Elena Voita, Rico Sennrich, and Ivan Titov.
\newblock Analyzing the source and target contributions to predictions in neural machine translation.
\newblock In \emph{Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)}, pages 1126--1140, 2021.
\bibitem[Xiong et~al.(2024)Xiong, Li, Zhang, Chen, Sun, Li, Sun, and Du]{xiong2024explainableartificialintelligencexai}
Haoyi Xiong, Xuhong Li, Xiaofei Zhang, Jiamin Chen, Xinhao Sun, Yuchen Li, Zeyi Sun, and Mengnan Du.
\newblock Towards explainable artificial intelligence (xai): A data mining perspective, 2024.
\newblock URL \url{https://arxiv.org/abs/2401.04374}.
\bibitem[Yuan et~al.(2021)Yuan, Li, Xiong, Cao, and Dou]{yuan2021explaining}
Tingyi Yuan, Xuhong Li, Haoyi Xiong, Hui Cao, and Dejing Dou.
\newblock Explaining information flow inside vision transformers using markov chain.
\newblock In \emph{eXplainable AI approaches for debugging and diagnosis.}, 2021.
\bibitem[Zhang et~al.(2024)Zhang, Zeng, Wang, and Lu]{zhang2024tinyllama}
Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu.
\newblock Tinyllama: An open-source small language model.
\newblock \emph{arXiv preprint arXiv:2401.02385}, 2024.
\bibitem[Zhang and Zhu(2018)]{zhang2018visual}
Quan-shi Zhang and Song-Chun Zhu.
\newblock Visual interpretability for deep learning: a survey.
\newblock \emph{Frontiers of Information Technology \& Electronic Engineering}, 19\penalty0 (1):\penalty0 27--39, 2018.
\bibitem[Zhang et~al.(2021)Zhang, Ti{\v{n}}o, Leonardis, and Tang]{zhang2021survey}
Yu~Zhang, Peter Ti{\v{n}}o, Ale{\v{s}} Leonardis, and Ke~Tang.
\newblock A survey on neural network interpretability.
\newblock \emph{IEEE Transactions on Emerging Topics in Computational Intelligence}, 5\penalty0 (5):\penalty0 726--742, 2021.
\bibitem[Zimerman et~al.(2025)Zimerman, Ali, and Wolf]{zimerman2025explaining}
Itamar Zimerman, Ameen~Ali Ali, and Lior Wolf.
\newblock Explaining modern gated-linear {RNN}s via a unified implicit attention formulation.
\newblock In \emph{The Thirteenth International Conference on Learning Representations}, 2025.
\newblock URL \url{https://openreview.net/forum?id=wnT8bfJCDx}.
\end{thebibliography}