Selective Capability Unlearning in End-to-End Spoken Language Understanding
Extracted source text · arXiv 2606.24063
% *==================================================================================*
% * Review vs. Camera-Ready settings *
% *==================================================================================*
%
% REVIEW: Use the following command for submitting the paper (double-blind,
% for review):
% \documentclass{Interspeech}
%
% CAMERA-READY: Use the following command for the camera-ready version, one
% affiliation per line:
\documentclass[cameraready]{Interspeech}
% *==================================================================================*
\newcommand{\blue}[1]{\textcolor{blue}{#1}}
\newcommand{\bcirc}[1]{\raisebox{0.5pt}{\textcircled{\footnotesize\bfseries #1}}}
% general
\usepackage{comment}
\usepackage{pifont}
% algorithm
\usepackage{algorithm}
\usepackage{algpseudocode}
\usepackage{amsmath,amssymb}
% tables
\usepackage{graphicx} % for \resizebox
\usepackage[table]{xcolor}
\setlength{\tabcolsep}{2pt} % even tighter % smaller font than \small
\usepackage{colortbl}
\usepackage{booktabs}
\usepackage{adjustbox}
\usepackage{multirow}
\usepackage{array}
% \usepackage[export]{adjustbox}
\usepackage{subcaption}
\usepackage{float}
\setlength{\textfloatsep}{6pt}
\setlength{\floatsep}{6pt}
% **************************************
% * *
% * STOP ! DO NOT DELETE ! *
% * READ THIS FIRST *
% * *
% * This template also includes *
% * important INSTRUCTIONS that you *
% * must follow when preparing your *
% * paper. Read it BEFORE replacing *
% * the content with your own work. *
% **************************************
%==================================================================================
% Title
% Must exactly match the title entered into the paper submission system
\title{Selective Capability Unlearning in End-to-End Spoken Language Understanding}
%==================================================================================
% Authors
% The order of authors here must exactly match the order entered into the paper submission system
% Note that the COMPLETE list of authors MUST be entered into the paper submission system at the outset, including when submitting your manuscript for double-blind review
% The ORCID number is still optional but will become mandatory in the future years. It is strongly encouraged to get an ORCID for each cu-author.
% Middle names, including initials, must be included in the first name
%==================================================================================
% Authors
\author[affiliation={1}, orcid=0009-0005-1624-272X]{Akanksha}{Singh}
\author[affiliation={1}, orcid=0000-0003-0332-5647]{Vinod Kumar}{Kurmi}
%==================================================================================
% Affiliations
\address{
$^1$ Indian Institute of Science Education and Research Bhopal, India
}
%==================================================================================
% Emails
\email{akanksha23@iiserb.ac.in, vinodkk@iiserb.ac.in}
%==================================================================================
% Keywords
\keywords{Machine Unlearning, Spoken Language Understanding, Speech Recognition.}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\usepackage{comment}
%==================================================================================
% Content
\begin{document}
\maketitle
\begin{abstract}
Modern spoken language understanding (SLU) systems are increasingly deployed in real-world settings, where specific functionalities may need to be removed due to policy or safety constraints. In SLU, a functionality corresponds to an intent and its associated slot-generation behavior. However, in autoregressive models, suppressing a target intent does not eliminate the conditional mapping that generates slots conditioned on that intent. When the intent prefix is externally supplied, the model can reconstruct the original intent-slot structure. We identify this structural failure as \textbf{\emph{capability persistence}}. We propose \textit{\underline{B}inding \underline{S}ubspace (BSU)}, a representation-level framework that isolates and attenuates intent-conditioned directions underlying this mapping. Across SLU benchmarks, BSU substantially reduces forced-prefix recoverability while preserving retained performance.
\end{abstract}
\section{Introduction}
Spoken language understanding (SLU) constitutes a core component of conversational systems. It enables devices like voice assistants and spoken interfaces to extract structured semantic information directly from speech \cite{bastianelli2020slurp, wang2021fine, arora2022espnet, seo2022integration}. Modern end-to-end SLU models \cite{huang2023leveraging, sharma2021leveraging} directly map acoustic input to semantic outputs and are widely adapted in diverse applications, including virtual assistants, customer support systems, in-vehicle voice control, and domain-specific voice agents \cite{ma2024speech, mehta2020recent}.
\vspace{0.2em}
As these systems are increasingly deployed in real-world and regulated settings, there is a growing need to adjust model behavior post-training, including SLU frameworks \cite{koudounas2025alexa}. In practice, certain functionalities may need to be disabled due to policy changes, safety considerations, or regulatory requirements \cite{voigt2017eu, goldman2020introduction, xu2024machine}. For example, a voice assistant may need to deactivate financial transaction capabilities in specific regions or restrict health-related guidance under updated compliance rules. Retraining these models from scratch in such cases is often costly and impractical, motivating the need for targeted post-deployment methods that can selectively remove undesired behaviors while preserving other functionalities.
\vspace{0.2em}
However, in an autoregressive SLU, functionalities are not merely intent labels. They correspond to an intent together with its associated slot-generation behavior. Semantic generation is inherently conditional, as the decoder first predicts an intent token, and subsequent slot tokens are generated conditioned on that prefix. Thus, the behavior associated with an intent is governed by the conditional mapping from acoustic input and intent prefix to slot values. In this work, we define the removable functionality as precisely this intent-conditioned conditional mapping. Upon a request to remove functionality, existing methods typically suppress the marginal probability of a target intent. However, this does not necessarily modify the conditional mapping responsible for slot generation. Consequently, under a forced intent prefix, the model can still reconstruct the corresponding intent-slot structure. We refer to this phenomenon as capability persistence. As illustrated in Figure~\ref{fig:01}, although intent prediction is suppressed under standard decoding, providing the intent as a prefix enables recovery of the associated slot structure. This limitation arises from the conditional generative structure of autoregressive SLU models, which classification-based unlearning methods do not explicitly address. To eliminate the conditional mapping itself, we propose a two-stage framework.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{Figures/1122.jpg}
\vspace{-0.45cm}
\caption{\textbf{Capability persistence under forced-prefix decoding.} Existing methods suppress marginal intent prediction but preserve conditional intent–slot binding, enabling reconstruction under forced prefix. BSU removes this dependency.}
% \vspace{-2em}
\label{fig:01}
\end{figure}
\vspace{0.3em}
\noindent\bcirc{1}\ \textbf{Binding Subspace Identification.}
For a target intent $I_F$, we identify representation directions that capture its slot-generation behavior. We extract decoder hidden states at slot positions for forget and retain data and compute layer-wise covariances. By contrasting these covariances, we extract the eigen-directions with the largest positive variance excess under $I_F$, forming a compact subspace that captures intent-slot dependency.
\noindent\bcirc{2}\ \textbf{Subspace-Guided Capability Attenuation.}
We fine-tune the model while reducing sensitivity along the identified subspace. Specifically, we apply gradient-based regularization to attenuate dependence on these directions, thereby weakening the conditional mapping while preserving general performance.
\vspace{0.3em}
Our key contributions are outlined as follows:
\textbf{\textit{(i)}} We formalize selective capability unlearning in SLU by defining intent capability as the conditional mapping and identifying \emph{capability persistence} as recoverable slot generation under forced-prefix decoding.
\textbf{\textit{(ii)}} We propose \textit{Binding Subspace Unlearning} (BSU)\footnote{Annotations and code will be made publicly available.}, a representation-level framework that localizes intent-conditioned binding directions via covariance contrast and attenuates them through subspace-guided gradient regularization.
\textbf{\textit{(iii)}} We introduce a recoverability-based evaluation protocol to measure residual conditional behavior beyond intent accuracy.
\textbf{\textit{(iv)}} BSU reduces conditional slot recoverability, with an average drop of $\sim$60\% in BRR@10 and $\sim$56\% in semantic similarity, while preserving retained-intent performance without inference-time overhead.
\section{Problem Formulation}
\vspace{-0.1cm}
We consider an end-to-end spoken language understanding (SLU) model that maps an input speech signal $x \in \mathcal{X}$ to a structured semantic frame represented as a token sequence $y=(y_1,\dots,y_T)$. The sequence follows a fixed format in which the initial tokens encode an intent label $i \in \mathcal{I}$, followed by slot-type and slot-value tokens $s \in \mathcal{S}$. The model is autoregressive and parameterized by $\theta$, defining
\begingroup
\setlength{\abovedisplayskip}{3pt}
\setlength{\belowdisplayskip}{3pt}
\setlength{\abovedisplayshortskip}{3pt}
\setlength{\belowdisplayshortskip}{3pt}
\[
p_{\theta}(y \mid x)
=
\prod_{t=1}^{T}
p_{\theta}(y_t \mid y_{<t},x).
\]
\endgroup
Because intent tokens precede slot tokens in the decoding order, the joint distribution factorizes as
\begingroup
\setlength{\abovedisplayskip}{4pt}
\setlength{\belowdisplayskip}{4pt}
\setlength{\abovedisplayshortskip}{4pt}
\setlength{\belowdisplayshortskip}{4pt}
\[
p_{\theta}(i,s|x)
=
p_{\theta}(i|x)\,
p_{\theta}(s|i,x),
\]
\endgroup
where slot generation is conditioned jointly on the acoustic input and the decoded intent prefix. We define the semantic capability associated with intent $i$ as the conditional mapping
\begingroup
\setlength{\abovedisplayskip}{4pt}
\setlength{\belowdisplayskip}{4pt}
\setlength{\abovedisplayshortskip}{4pt}
\setlength{\belowdisplayshortskip}{4pt}
\[
C_{\theta}(i)(x)=p_{\theta}(s|i,x),
\]
\endgroup
that characterizes the model’s ability to generate slot values given the intent and acoustic input. This definition distinguishes capability from marginal intent prediction. Suppressing the marginal probability $p_{\theta}(i|x)$ does not necessarily eliminate the conditional mapping $p_{\theta}(s|i,x)$ when the intent prefix is externally supplied during decoding. This motivates the problem of \textit{selective capability unlearning}.
\vspace{-1em}
\subsection{Selective Capability Unlearning }
\vspace{-0.5em}
Let $I_F \in \mathcal{I}$ denote a target intent whose associated capability must be removed. The training corpus is partitioned into a forget set $\mathcal{D}_F$, containing samples labeled with $I_F$, and a retain set $\mathcal{D}_R$, containing all remaining intents. Given pretrained parameters $\theta_0$, the objective is to obtain parameters $\theta^*$ such that:
\begin{enumerate}[label=\ding{\numexpr171+\arabic*}, leftmargin=2em]
\item \textbf{Capability Erasure}. The conditional mapping $p_{\theta^*}(s \mid i_f, x)$ must be substantially degraded for inputs $x \sim \mathcal{D}_F$, such that even when the intent prefix $I_F$ is externally supplied, the model can no longer generate correct slots.
\item \textbf{Capability Retention}. For all $i \neq i_f$, the joint behavior $p_{\theta^*}(i, s \mid x)$ should remain close to that under $\theta_0$, thereby preserving performance on all retained intents.
\end{enumerate}
In other words, our objective is to reduce the conditional mapping associated with the target intent while preserving the remaining capabilities of the model.
\vspace{-0.1cm}
\subsection{Failure Mode: Capability Persistence}
Existing unlearning methods typically suppress the marginal likelihood $p_{\theta}(i_f \mid x)$ of the target intent. However, this does not
necessarily modify the conditional slot-generation distribution
$p_{\theta}(s \mid i_f, x)$. This limitation arises from the autoregressive decoding. For slot positions
$t > t_{\text{intent}}$, the decoder hidden state $h_t = f_{\theta}(x, i_f, s_{<t})$
depends jointly on the acoustic input $x$, the intent prefix $I_F$, and
previously generated tokens. The next-token probability is computed as
% \vspace{-0.2cm}
\begingroup
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\abovedisplayshortskip}{1pt}
\setlength{\belowdisplayshortskip}{2pt}
\[
p_{\theta}(y_t \mid y_{<t}, x) = \mathrm{softmax}(W h_t + b).
\]
\endgroup
Here, $f_{\theta}$ denotes the decoder network parameterized by $\theta$, while $W$ and $b$ are the output projection parameters mapping hidden states to vocabulary logits. Because representation directions in $h_t$ encode the intent-slot associations learned during training, suppressing $p_{\theta}(i_f \mid x)$ does not necessarily modify these directions. As shown in Table~\ref{tab:02}, intent-level suppression reduces intent accuracy but leaves conditional slot recoverability largely intact. Consequently, the conditional mapping $p_{\theta}(s \mid i_f, x)$ remains active, and correct slots can still be generated under forced-prefix decoding. We refer to this structural phenomenon as \emph{capability persistence}. Effective selective unlearning should therefore intervene in the representation space to weaken intent-conditioned slot generation rather than only suppress marginal intent prediction.
\section{BSU: Binding Subspace Unlearning}
Building on this formulation, we introduce \textit{Binding Subspace Unlearning (BSU)}, a two-stage unlearning framework that intervenes in the model's representation space to mitigate capability persistence. Rather than suppressing only the marginal intent probability $p_{\theta}(i_f \mid x)$, BSU targets hidden-state directions associated with intent-conditioned slot generation. Let $\theta_0$ be a pretrained SLU model and $\theta$ a trainable copy initialized from $\theta_0$. Given a target intent $I_F$, the objective is to reduce the conditional distribution $p_{\theta}(s \mid i_f, x)$ under forced-prefix decoding while preserving performance on retained intents. BSU achieves this through two stages: \textit{\textbf{(i)}} identifying representation directions enriched when the target intent $I_F$ is present, and \textit{\textbf{(ii)}} reducing model sensitivity along these directions to weaken the conditional mapping.
\vspace{-0.2em}
\subsection{Stage I: Binding Subspace Identification}
\label{sec:subspace_id}
\vspace{-0.2cm}
The goal of Stage I is to identify representation directions that are statistically enriched when the target intent $I_F$ is present. These directions approximate components of the hidden state that influence the conditional slot-generation distribution $p_\theta(s \mid i_f, x)$. Since next-token probabilities are computed from hidden states, analyzing hidden representations at slot positions allows us to probe directions associated with intent-conditioned slot generation.
\noindent\textit{\underline{(1) Hidden-State Extraction. }} Semantic outputs follow a structured prefix format consisting of an intent token followed by slot tokens. We estimate the boundary between intent and slot tokens and extract hidden states at slot positions using teacher-forced decoding \cite{williams1989learning}:
\begingroup
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\abovedisplayshortskip}{2pt}
\setlength{\belowdisplayshortskip}{2pt}
\[
h_t^{(\ell)} = f_{\theta}^{(\ell)}(x, y_{<t}),
\qquad
t > t_{\text{prefix}}.
\]
\endgroup
Here, $h_t^{(\ell)}$ denotes the decoder representation at layer $\ell$. Teacher forcing ensures alignment between hidden states and ground truth slot structure, avoiding variability induced by decoding errors.
\vspace{0.3em}
\noindent\textit{\underline{(2) Covariance Contrast. }}Using these slot-position representations $h_t^{(\ell)}$, we compute empirical covariance matrices over hidden states for both the forget dataset $\mathcal{D}_F$ and the retain dataset $\mathcal{D}_R$ at each layer $\ell$. We then form the contrast matrix as:
\begingroup
\setlength{\abovedisplayskip}{4pt}
\setlength{\belowdisplayskip}{4pt}
\setlength{\abovedisplayshortskip}{4pt}
\setlength{\belowdisplayshortskip}{4pt}
\[
M^{(\ell)}
=
\mathrm{Cov}_{\mathcal{D}_F}^{(\ell)}
-
\mathrm{Cov}_{\mathcal{D}_R}^{(\ell)}
\]
\endgroup
This contrast highlights representation directions that have higher variance when the target intent $I_F$ is present relative to retained data. Since next-token probabilities are computed as affine transformations of hidden states, directions that consistently vary under $D_F$ are expected to contribute more strongly to the conditional slot likelihood $p_\theta(s \mid i_f, x)$.
\vspace{0.4em}
\noindent\textit{\underline{(3) Subspace Extraction. }}Finally, we compute the top positive eigenvectors of $M^{(\ell)}$, denoted as $U^{(\ell)} \in \mathbb{R}^{d \times k}$. These eigenvectors define a low-dimensional subspace that approximates representation directions associated with intent-conditioned slot generation. This construction is statistical rather than exact; it identifies directions enriched under $I_F$ without assuming strict disentanglement of representations.
\subsection{Stage II: Subspace-Guided Capability Attenuation}
\label{sec:subspace_att}
Stage~II attenuates intent-conditioned slot generation by reducing model sensitivity along the identified binding subspaces $U^{(\ell)}$. Rather than modifying only output distributions, BSU targets the first-order sensitivity of the conditional likelihood with respect to hidden states. For slot positions associated with $I_F$, we compute the gradient of the teacher-forced conditional log-likelihood:
\begingroup
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\abovedisplayshortskip}{2pt}
\setlength{\belowdisplayshortskip}{2pt}
\[
g_t^{(\ell)}=\nabla_{h_t^{(\ell)}}\log p_\theta(s\mid i_f,x)
\]
\endgroup
If slot generation under $I_F$ relies on particular representation directions, the gradient will exhibit large components along them. We therefore project the gradients onto the binding subspaces and penalize their squared magnitude:
\begingroup
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\abovedisplayshortskip}{2pt}
\setlength{\belowdisplayshortskip}{2pt}
\[
\mathcal{L}_{\mathrm{bind}}
=
\sum_{\ell=1}^{L}
\sum_{t>t_{\mathrm{prefix}}}
\left\lVert
U^{(\ell)}U^{(\ell)\top}g_t^{(\ell)}
\right\rVert_2^2
\]
\endgroup
Minimizing $\mathcal{L}_{\mathrm{bind}}$ reduces sensitivity along representation directions enriched under $I_F$, weakening intent-conditioned slot generation even when the intent prefix is externally supplied. The attenuation is achieved through parameter updates and introduces no inference-time overhead.
% \subsection{Overall Training Objective}
% The training objective combines standard forget--retain fine-tuning with the proposed binding-sensitivity regularizer. We define the negative log-likelihood objective $L_{\text{NLL}}(D)=-\mathbb{E}_{(x,y)\sim D}\log p_\theta(y\mid x)$.
% The forget and retain losses correspond to
% \begingroup
% \setlength{\abovedisplayskip}{2pt}
% \setlength{\belowdisplayskip}{2pt}
% \setlength{\abovedisplayshortskip}{2pt}
% \setlength{\belowdisplayshortskip}{2pt}
% \[
% L_{\text{forget}}=L_{\text{NLL}}(D_F), \qquad
% L_{\text{ret}}=L_{\text{NLL}}(D_R).
% \]
% \endgroup
% The forget loss suppresses the target capability on $D_F$, while the retain loss preserves performance on non-target intents in $D_R$. To stabilize optimization and limit drift from the original model $\theta_0$, we additionally include KL regularization.
% \begingroup
% \setlength{\abovedisplayskip}{2pt}
% \setlength{\belowdisplayskip}{2pt}
% \setlength{\abovedisplayshortskip}{2pt}
% \setlength{\belowdisplayshortskip}{2pt}
% \[
% L_{\text{kl}}=\mathbb{E}_{x\sim D_R}\!\left[
% \mathrm{KL}\big(p_{\theta_0}(\cdot\mid x)\,\|\,p_\theta(\cdot\mid x)\big)
% \right].
% \]
% \endgroup
% Finally, we incorporate the binding loss from Section~3.2 to attenuate representation directions associated with intent-conditioned slot generation. The overall objective is
% \begingroup
% \setlength{\abovedisplayskip}{2pt}
% \setlength{\belowdisplayskip}{2pt}
% \setlength{\abovedisplayshortskip}{2pt}
% \setlength{\belowdisplayshortskip}{2pt}
% \[
% L=L_{\text{forget}}
% +\lambda_{\text{ret}}L_{\text{ret}}
% +\lambda_{\text{kl}}L_{\text{kl}}
% +\lambda_{\text{bind}}L_{\text{bind}} .
% \]
% \endgroup
% Here $\lambda_{\text{ret}},\lambda_{\text{kl}}$, and $\lambda_{\text{bind}}$ control the relative contribution of each term.
% % \subsection{Overall Training Objective. } The training objective combines standard forget--retain fine-tuning with the proposed binding-sensitivity regularizer. We first define the negative log-likelihood objective:$
% % L_{\text{NLL}}(D) = -\mathbb{E}_{(x,y)\sim D}\log p_\theta(y\mid x)$. Using this formulation, the forget and retain losses correspond to
% % \vspace{-0.1cm}
% % \begingroup
% % \setlength{\abovedisplayskip}{2pt}
% % \setlength{\belowdisplayskip}{2pt}
% % \setlength{\abovedisplayshortskip}{2pt}
% % \setlength{\belowdisplayshortskip}{2pt}
% % \[
% % L_{\text{forget}} = L_{\text{NLL}}(D_F), \qquad
% % L_{\text{ret}} = L_{\text{NLL}}(D_R).
% % \]
% % \endgroup
% % The forget loss suppresses the target capability on $D_F$, while the retain loss preserves model performance on non-target intents in $D_R$. To stabilize the unlearning objective and prevent excessive deviation from the original model parameters $\theta_0$, we additionally include a KL regularization term:
% % \begingroup
% % \setlength{\abovedisplayskip}{3pt}
% % \setlength{\belowdisplayskip}{3pt}
% % \setlength{\abovedisplayshortskip}{3pt}
% % \setlength{\belowdisplayshortskip}{3pt}
% % \[
% % L_{\text{kl}}
% % =
% % \mathbb{E}_{x \sim D_R}\!\left[
% % \mathrm{KL}\!\left(
% % p_{\theta_0}(\,\cdot \mid x)\,\|\,p_{\theta}(\,\cdot \mid x)
% % \right)
% % \right]
% % \]
% % \endgroup
% % Finally, to directly address the representation-level mechanism underlying capability persistence, we incorporate the \textit{binding loss} introduced in Section~3.2, which attenuates gradients along representation directions associated with intent-conditioned slot generation. The overall training objective is therefore
% % \begingroup
% % \setlength{\abovedisplayskip}{2pt}
% % \setlength{\belowdisplayskip}{2pt}
% % \[
% % L =
% % L_{\text{forget}}
% % + \lambda_{\text{ret}} L_{\text{ret}}
% % + \lambda_{\text{kl}} L_{\text{kl}}
% % + \lambda_{\text{bind}} L_{\text{bind}}
% % \]
% % \endgroup
% % where $\lambda_{\text{ret}}$, $\lambda_{\text{kl}}$, and $\lambda_{\text{bind}}$ control the relative contribution of each objective component.
% % --------------------------
% % \section{ BSU: Binding Subspace Unlearning}
% % To address this limitation, we now introduce \textit{BSU: Binding Subspace Unlearning} as a two-stage, representation-level unlearning framework. First, in section 3.1, we describe \emph{Binding Subspace Identification}, which isolates hidden-state directions associated with the target intent. Then in section 3.2, we present \emph{Subspace-Guided Gradient Attenuation}, where we apply gradient-based regularization to attenuate sensitivity to the identified directions, thereby suppressing intent-conditioned slot generation. By selectively intervening on representation directions that support intent-conditioned slot generation, BSU targets capability persistence directly rather than suppressing marginal intent prediction.
% % \subsection{Stage I: Binding Subspace Identification}
% % Stage~I aims to identify representation directions associated with intent-conditioned slot generation under the target intent $I_F$. Because hidden states parameterize next-token probabilities through affine projections to the vocabulary space, directions that systematically modulate the conditional likelihood $p_{\theta}(s \mid i_f, x)$ are expected to exhibit distinctive activation statistics at slot-token positions. We therefore seek directions whose variance is selectively enriched on $\mathcal{D}_F$ relative to retained data $\mathcal{D}_R$.
% % \noindent We exploit the structured output format in which intent tokens precede slot tokens. Let $t_{\text{prefix}}$ denote the estimated boundary separating intent and slot positions. For each layer $\ell$, we extract hidden states at slot positions under teacher-forced decoding:
% % \begin{equation}
% % h_t^{(\ell)} = f_{\theta}^{(\ell)}(x, y_{<t}),
% % \quad t > t_{\text{prefix}} .
% % \end{equation}
% % Teacher forcing preserves alignment with the ground-truth slot structure and prevents variability induced by autoregressive decoding errors.For each layer $\ell$, we compute empirical covariance matrices over slot-position hidden states on $\mathcal{D}_F$ and $\mathcal{D}_R$, respectively, and define the contrast matrix
% % \begin{equation}
% % M^{(\ell)}
% % =
% % \mathrm{Cov}_{\mathcal{D}_F}^{(\ell)}
% % -
% % \mathrm{Cov}_{\mathcal{D}_R}^{(\ell)} .
% % \end{equation}
% % Since output logits are linear transformations of hidden states, directions exhibiting elevated variance under $\mathcal{D}_F$ are hypothesized to correspond to components that more strongly influence the conditional slot distribution under $I_F$. We obtain a low-dimensional approximation of this subspace by computing the top-$k$ eigenvectors of $M^{(\ell)}$, denoted $U^{(\ell)} \in \mathbb{R}^{d \times k}$. These eigenvectors define a statistical estimate of representation directions enriched under $I_F$. This construction does not assume strict disentanglement; rather, it identifies dominant variance components associated with intent-conditioned slot generation.
% % \begin{algorithm}[t]
% % \caption{Two-Stage Capability Unlearning \textit{(Ours)}}
% % \label{alg:ours_compact}
% % \centering
% % \begingroup
% % \footnotesize
% % \resizebox{\linewidth}{!}{%
% % \begin{minipage}{\linewidth}
% % \begin{algorithmic}[1]
% % \Require Pretrained parameters $\theta_0$; forget set $\mathcal{D}_F$ (intent $I_F$); retain set $\mathcal{D}_R$; decoder layers $\ell=1,\dots,L$; rank $k$; weights $\lambda_{\mathrm{ret}},\lambda_{\mathrm{kl}},\lambda_{\mathrm{bind}}$; steps $E$; learning rate $\eta$
% % \Ensure Unlearned parameters $\theta^\star$
% % \Statex \textbf{Stage I: Binding subspace identification.}
% % \State $\theta \gets \theta_0$
% % \For{$\ell = 1$ to $L$}
% % \State Collect teacher-forced slot states $\{h_t^{(\ell)}\}_{t>t_{\mathrm{prefix}}}$ on $\mathcal{D}_F$ and $\mathcal{D}_R$
% % \State $\Sigma_F^{(\ell)} \gets \mathrm{Cov}_{(x,y)\sim\mathcal{D}_F}\!\left[h_t^{(\ell)}\right]$
% % \State $\Sigma_R^{(\ell)} \gets \mathrm{Cov}_{(x,y)\sim\mathcal{D}_R}\!\left[h_t^{(\ell)}\right]$
% % \State $M^{(\ell)} \gets \Sigma_F^{(\ell)} - \Sigma_R^{(\ell)}$
% % \State $U^{(\ell)} \gets \mathrm{TopEig}(M^{(\ell)}, k)$ \Comment{$U^{(\ell)}\in\mathbb{R}^{d\times k}$}
% % \EndFor
% % \Statex \textbf{Stage II: Subspace-guided attenuation.}
% % \For{step $=1$ to $E$}
% % \State Sample mixed minibatch $(B_F, B_R)$ from $\mathcal{D}_F, \mathcal{D}_R$
% % \State $\mathcal{L}_{\mathrm{forget}} \gets -\log p_\theta(y\mid x)$ on $B_F$ \Comment{GA: maximize loss on $B_F$}
% % \State $\mathcal{L}_{\mathrm{ret}} \gets -\log p_\theta(y\mid x)$ on $B_R$
% % \State $\mathcal{L}_{\mathrm{kl}} \gets \mathrm{KL}\!\left(p_{\theta_0}(\cdot\mid x)\,\Vert\,p_\theta(\cdot\mid x)\right)$ on $B_R$
% % \State $\mathcal{L}_{\mathrm{bind}} \gets \sum_{(x,y)\in B_F}\sum_{\ell=1}^{L}\sum_{t>t_{\mathrm{prefix}}}
% % \left\lVert U^{(\ell)}U^{(\ell)\top} g_t^{(\ell)} \right\rVert_2^2$
% % \Statex \hspace{2.0em} where $g_t^{(\ell)}=\nabla_{h_t^{(\ell)}}\log p_\theta\!\left(y_t \mid y_{<t}, x\right)$ (teacher forcing)
% % \State $\mathcal{L} \gets \mathcal{L}_{\mathrm{forget}} + \lambda_{\mathrm{ret}}\mathcal{L}_{\mathrm{ret}} + \lambda_{\mathrm{kl}}\mathcal{L}_{\mathrm{kl}} + \lambda_{\mathrm{bind}}\mathcal{L}_{\mathrm{bind}}$
% % \State $\theta \gets \theta - \eta \nabla_\theta \mathcal{L}$
% % \EndFor
% % \State \Return $\theta^\star \gets \theta$
% % \end{algorithmic}
% % \end{minipage}%
% % } % <-- closes \resizebox third argument
% % \endgroup
% % \end{algorithm}
\begingroup
\scriptsize
\setlength{\tabcolsep}{1.6pt}
\renewcommand{\arraystretch}{0.85}
\setlength{\extrarowheight}{0.8pt}
\begin{table*}[!t]
\centering
\vspace{-1cm}
\caption{Selective capability unlearning on \textbf{SLURP} and \textbf{SpeechMassive}. We report intent accuracy ($I_F$), slot macro-F1 ($F1_F$), BRR@10, and semantic similarity (Sim.) on the target $D_F$ and retain $D_R$ sets, respectively. \textbf{Bold}: Best, \underline{Underlined:} Second-Best.}
\label{tab:02}
\vspace{-0.3cm}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{|l|l|cccc|cccc||cccc|cccc|}
\toprule
\multirow{3}{*}{\textbf{Dataset}}
& \multirow{3}{*}{\textbf{Method}}
& \multicolumn{8}{c||}{\textbf{NeMo-Conformer-Transformer}}
& \multicolumn{8}{c}{\textbf{NeMo-SSL-Conformer}} \\
\cmidrule(lr){3-10}\cmidrule(lr){11-18}
&
& \multicolumn{4}{c|}{\textbf{Forget Set ($D_F$)} }
& \multicolumn{4}{c||}{\textbf{Retain Set ($D_R$)}}
& \multicolumn{4}{c|}{\textbf{Forget Set ($D_F$)}}
& \multicolumn{4}{c}{\textbf{Retain Set ($D_R$)}} \\
\cmidrule(lr){3-6}\cmidrule(lr){7-10}\cmidrule(lr){11-14}\cmidrule(lr){15-18}
&
& $I_F\,\downarrow$ & $F1_F\,\downarrow$ & $\mathrm{BRR}@10\,\downarrow$ & $\mathrm{Sim.}\,\downarrow$
& $I_F\,\uparrow$ & $F1_F\,\uparrow$ & $\mathrm{BRR}@10\,\uparrow$ & $\mathrm{Sim.}\,\uparrow$
& $I_F\,\downarrow$ & $F1_F\,\downarrow$ & $\mathrm{BRR}@10\,\downarrow$ & $\mathrm{Sim.}\,\downarrow$
& $I_F\,\uparrow$ & $F1_F\,\uparrow$ & $\mathrm{BRR}@10\,\uparrow$ & $\mathrm{Sim.}\,\uparrow$ \\
\midrule
% ================= SLURP =================
\multirow{8}{*}{\rotatebox[origin=c]{90}{\textbf{SLURP}}}
& \cellcolor{gray!23} Original
& \cellcolor{gray!23}95.27 & \cellcolor{gray!23}91.08 & \cellcolor{gray!23}92.64 & \cellcolor{gray!23}90.14
& \cellcolor{gray!23}89.75 & \cellcolor{gray!23}80.11 & \cellcolor{gray!23}91.17 & \cellcolor{gray!23}90.11
& \cellcolor{gray!23}91.05 & \cellcolor{gray!23}84.90 & \cellcolor{gray!23}84.42 & \cellcolor{gray!23}83.27
& \cellcolor{gray!23}89.16 & \cellcolor{gray!23}78.90 & \cellcolor{gray!23}85.17 & \cellcolor{gray!23}86.32 \\
& \cellcolor{yellow!20} Retrain
& \cellcolor{yellow!20}14.04 &\cellcolor{yellow!20} 20.13&\cellcolor{yellow!20} 18.59&\cellcolor{yellow!20}23.53
& \cellcolor{yellow!20}90.14 & \cellcolor{yellow!20}86.58 &\cellcolor{yellow!20}81.38 &\cellcolor{yellow!20}82.10
& \cellcolor{yellow!20}13.72 & \cellcolor{yellow!20}16.58 & \cellcolor{yellow!20}21.38 & \cellcolor{yellow!20}22.19
& \cellcolor{yellow!20}90.15 &\cellcolor{yellow!20}80.43 & \cellcolor{yellow!20}81.70 &\cellcolor{yellow!20}83.27 \\
& GA & \textbf{21.36} & \underline{31.78} & 91.14 & 86.92 & 54.15 & 52.09 & 87.62 & 83.40
& \underline{14.80} & \underline{22.50} & \underline{63.10} & \underline{66.40}
& 89.16 & 78.00 & 85.42 & 84.27 \\
& GA+GD & 63.23 & 69.50 & 94.05 & 89.90 & 85.57 & 84.16 & 83.06 & 87.42
& 48.60 & 54.20 & 71.90 & 74.13
& 85.05 & 79.43 & 85.01 & 88.76 \\
& GA+KL & 42.60 & 44.40 & 94.78 & 88.73 & 81.49 & 79.04 & 88.17 & 86.84
& 28.90 & 31.70 & 69.40 & 71.00
& 80.11 & 78.39 & 81.07 & 81.72 \\
& NPO & 54.85 & 52.73 & 87.62 & 86.78 & 84.05 & 82.21 & 82.14 & 87.91
& 36.20 & 39.80 & 61.50 & 65.90
& 88.25 & 79.00 & 80.11 & 86.80 \\
& NPO+KL & 84.17 & 82.62 & 84.05 & 82.79 & 89.68 & 82.94 & 85.07 & 80.17
& 62.70 & 58.10 & 79.81 & 76.20
& 89.01 & 80.51 & 83.30 & 89.05 \\
& RL &58.7 & 63.41 &71.60 &80.32 &88.91 &81.88 &84.72 & 85.50
& 61.5 &66.28 & 75.30 & 81.00
& 90.01 &79.26 & 81.00 & 83.49 \\
\cline{2-18}
& \textcolor{gray}{\textbf{\textit{RS (Ablation)}}} & \textcolor{gray}{62.70} & \textcolor{gray}{71.59} & \textcolor{gray}{87.01} & \textcolor{gray}{88.0} & \textcolor{gray}{82.53} & \textcolor{gray}{82.01} & \textcolor{gray}{90.14} & \textcolor{gray}{89.01} & \textcolor{gray}{50.15} & \textcolor{gray}{50.59} & \textcolor{gray}{69.11} & \textcolor{gray}{70.16} & \textcolor{gray}{90.18} & \textcolor{gray}{73.87} & \textcolor{gray}{80.10} & \textcolor{gray}{83.15} \\
& \cellcolor{blue!12}\textbf{\textit{BSU (Ours)}}
& \cellcolor{blue!12}\underline{28.40} & \cellcolor{blue!12}\textbf{16.22} & \cellcolor{blue!12}\textbf{22.10} & \cellcolor{blue!12}\textbf{24.80}
& \cellcolor{blue!12}87.90 & \cellcolor{blue!12}83.62 & \cellcolor{blue!12}88.53 & \cellcolor{blue!12}90.8
& \cellcolor{blue!12}\textbf{11.20} & \cellcolor{blue!12}\textbf{14.10} & \cellcolor{blue!12}\textbf{16.30} & \cellcolor{blue!12}\textbf{18.70}
& \cellcolor{blue!12}87.40 & \cellcolor{blue!12}79.14 & \cellcolor{blue!12}83.90 & \cellcolor{blue!12}86.16\\
\midrule \midrule
% ================= SpeechMassive =================
\multirow{8}{*}{\rotatebox[origin=c]{90}{\shortstack{\textbf{SpeechMassive}}}}
& \cellcolor{gray!23} Original
& \cellcolor{gray!23}81.55 &\cellcolor{gray!23}54.90 & \cellcolor{gray!23}79.02 & \cellcolor{gray!23}76.39
& \cellcolor{gray!23} 79.54 &\cellcolor{gray!23} 52.60 & \cellcolor{gray!23}78.49 & \cellcolor{gray!23} 77.83
& \cellcolor{gray!23} 77.38 &\cellcolor{gray!23}52.03 & \cellcolor{gray!23}77.02 & \cellcolor{gray!23} 74.70
& \cellcolor{gray!23} 73.66 &\cellcolor{gray!23} 51.09 & \cellcolor{gray!23} 76.02 & \cellcolor{gray!23}74.11 \\
& \cellcolor{yellow!20} Retrain
& \cellcolor{yellow!20} 12.74 & \cellcolor{yellow!20}23.01 &\cellcolor{yellow!20}19.77 &\cellcolor{yellow!20}21.80
& \cellcolor{yellow!20}80.52 & \cellcolor{yellow!20}54.11 & \cellcolor{yellow!20}80.14 & \cellcolor{yellow!20}79.44
& \cellcolor{yellow!20} 10.95 &\cellcolor{yellow!20}20.14 &\cellcolor{yellow!20}22.57 &\cellcolor{yellow!20}21.87
& \cellcolor{yellow!20}74.66 & \cellcolor{yellow!20}51.63 & \cellcolor{yellow!20}77.83 & \cellcolor{yellow!20}76.81 \\
& GA &33.15 &37.91 &31.77 &32.16 &48.66 &41.05 &45.92 &47.37 & 40.14 &45.80 &40.57 &44.60
&51.22 &46.03 &51.52 &50.87 \\
& GA+GD &77.51 &54.14 &78.01 &76.05 &78.05 &50.11 &68.50 &70.59 &74.06 &50.17 &67.55 &67.99
&72.19 &49.17 &68.05 &67.00 \\
& GA+KL &67.30 &48.01 &68.22 &71.00 &75.90 &53.80 &70.11 &71.08 &72.20 &48.68 &74.05 &74.70 &73.22 & 51.33 &73.59 &74.60 \\
& NPO &39.56 &44.17 &35.91 &35.07 &70.86 &51.73 &66.24 &67.38
&50.19 &50.11 &49.83 &49.17 &72.94 &53.16 &69.88 &70.42 \\
& NPO+KL &55.14 &49.76 &56.87 &58.43 &74.12 &53.92 &70.58 &71.16
&59.73 &51.34 &60.27 &61.43 &75.02 &55.94 &71.46 &72.08 \\
& RL &28.63 &36.42 &\textbf{24.94} &\textbf{26.18} &68.74 &50.91 &64.07 &65.32
&\underline{34.76} &\underline{41.27} &\underline{30.46} &\underline{32.11} &70.38 &52.48 &66.29 &67.44 \\
\cline{2-18}
& \textcolor{gray}{\textbf{\textit{RS (Ablation)}}}
& \textcolor{gray}{51.07} & \textcolor{gray}{47.50} & \textcolor{gray}{57.39} & \textcolor{gray}{59.47}
& \textcolor{gray}{67.92} & \textcolor{gray}{49.88} & \textcolor{gray}{63.15} & \textcolor{gray}{64.29}
& \textcolor{gray}{53.74} & \textcolor{gray}{45.03} & \textcolor{gray}{54.16} & \textcolor{gray}{51.97}
& \textcolor{gray}{69.81} & \textcolor{gray}{51.44} & \textcolor{gray}{65.88} & \textcolor{gray}{66.73} \\
& \cellcolor{blue!12}\textbf{\textit{BSU (Ours)}}
& \cellcolor{blue!12}\textbf{22.41} & \cellcolor{blue!12}\textbf{24.33 } & \cellcolor{blue!12}\underline{29.87} & \cellcolor{blue!12}\underline{29.94}
& \cellcolor{blue!12}79.14 & \cellcolor{blue!12}56.47 & \cellcolor{blue!12}78.36 & \cellcolor{blue!12}77.53
& \cellcolor{blue!12}\textbf{25.37} & \cellcolor{blue!12}\textbf{27.48} & \cellcolor{blue!12}\textbf{23.64} & \cellcolor{blue!12}\textbf{26.17}
& \cellcolor{blue!12}82.73 & \cellcolor{blue!12}58.26 & \cellcolor{blue!12}80.14 & \cellcolor{blue!12}79.47 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-2em}
\end{table*}
\endgroup
\subsection{Overall Unlearning Objective}
\label{sec:overall_obj}
\vspace{-0.15cm}
% For the final update step, we build on the standard forget-retain fine-tuning objective \cite{ren2025sok, kurmanji2023towards}. The forget loss $L_{DF} = -\mathbb{E}_{(x,y)\sim D_F}\log p_\theta(y|x)$ suppresses the target capability on the forget set $D_F$, while the retain loss $L_{DR} = -\mathbb{E}_{(x,y)\sim D_R}\log p_\theta(y|x)$ preserves performance on non-target intents in $D_R$. The base unlearning objective, therefore, combines these two components as follows:
% \begingroup
% \setlength{\abovedisplayskip}{4pt}
% \setlength{\belowdisplayskip}{4pt}
% \setlength{\abovedisplayshortskip}{4pt}
% \setlength{\belowdisplayshortskip}{4pt}
% \[
% L = L_{DF} + \lambda_{\text{ret}}\,L_{DR}
% \]
% \endgroup
% To stabilize optimization and limit deviation from the original model $\theta_0$, we additionally introduce a KL regularization term $L_{\text{kl}} =\mathbb{E}_{x\sim D_R}\mathrm{KL}\!\left(p_{\theta_0}(\cdot|x)\,\|\,p_\theta(\cdot|x)\right)$. This term constrains the updated model to remain close to the original distribution on retain set. Next, we incorporate the binding loss $L_{\text{bind}}$ from Section~3.2, which reduces model sensitivity along representation directions associated with intent-conditioned slot generation. The final training objective becomes:
% \vspace{0.4em}
% \begingroup
% \setlength{\abovedisplayskip}{4pt}
% \setlength{\belowdisplayskip}{4pt}
% \setlength{\abovedisplayshortskip}{4pt}
% \setlength{\belowdisplayshortskip}{4pt}
% \[
% L = L_{DF} + \lambda_{\text{ret}}L_{DR} + \lambda_{\text{kl}}L_{\text{kl}} + \lambda_{\text{bind}}L_{\text{bind}}
% \]
% \endgroup
For the final update step, we build on the standard forget--retain fine-tuning objective \cite{ren2025sok,kurmanji2023towards}. Let $L_F=-\mathbb{E}_{(x,y)\sim D_F}\log p_\theta(y\mid x)$ and $L_R=-\mathbb{E}_{(x,y)\sim D_R}\log p_\theta(y\mid x)$ denote the negative log-likelihood losses on the forget and retain sets, respectively. Since minimizing $L_F$ would further fit the forget examples, we reverse its sign in the unlearning objective. Thus, gradient descent on the total objective performs gradient ascent on $L_F$, suppressing the target capability, while $L_R$ preserves performance on non-target intents:
\[
\small
L_{\mathrm{base}}=-L_F+\lambda_{\mathrm{ret}}L_R .
\]
To stabilize optimization and limit deviation from the original model $\theta_0$, we introduce a retain-set KL regularizer $L_{\mathrm{kl}}$, which keeps the updated model close to the original distribution on $D_R$. We then incorporate the binding loss $L_{\mathrm{bind}}$ from Section~3.2 to reduce sensitivity along intent-conditioned slot-generation directions. The final objective is
\[
\small
L=-L_F+\lambda_{\mathrm{ret}}L_R+\lambda_{\mathrm{kl}}L_{\mathrm{kl}}+\lambda_{\mathrm{bind}}L_{\mathrm{bind}} .
\]
All terms are optimized with standard gradient descent; the negative coefficient on $L_F$ induces ascent on the forget-set NLL, while the retain, KL, and binding terms are minimized normally.
\section{Experimental Setup}
\subsection{ Datasets and Task Setup}
% \vspace{0.2em}
\noindent \textbf{Datasets. }We evaluate selective capability unlearning on SLURP \cite{bastianelli2020slurp}, a standard end-to-end SLU benchmark dataset with each utterance annotated by an intent and corresponding slot-value pairs. To assess cross-lingual robustness, we additionally evaluate the French subset of SpeechMASSIVE \cite{lee2024speech}, which follows the same semantic mapping. For each dataset, we select a target intent $I_F$ and partition the data into a forget set $D_F$ (utterances labeled with $I_F$) and a retain set $D_R$ (remaining intents).
\vspace{0.3em}
\noindent \textbf{Implementation Details. }All experiments use an end-to-end SLU architecture consisting of a Conformer acoustic encoder and a Transformer-based semantic decoder. We evaluate two model variants that share the same architecture, tokenizer, decoder, and training protocol, differing only in encoder initialization. In the \textit{(i) ASR-initialized SLU} setting, the encoder is initialized from a supervised ASR checkpoint (Conformer--Transformer--Large, NeMo ASR-Set~3.0). In the \textit{(ii) SSL-initialized SLU} setting, the encoder is instead initialized from a self-supervised speech representation while keeping all other components identical. Unless otherwise stated, we use $\lambda_{\mathrm{ret}}=1.0$, $\lambda_{\mathrm{kl}}=0.1$, and $\lambda_{\mathrm{bind}}=0.5$.
\subsection{Evaluation Metrics}
\vspace{-0.1cm}
We evaluate selective capability unlearning using both surface-level and behavioral metrics. Surface metrics include intent accuracy and slot F1 following the SLURP protocol \cite{bastianelli2020slurp}. To measure residual capability beyond exact prediction, we adopt top-k decoding and embedding-based similarity, widely used metrics in generative unlearning \cite{chen2021evaluating,zhang2019bertscore,yao2024large,carlini2021extracting,holtzman2019curious}.
\noindent \textbf{BRR@10 (Beam Retrieval Rate \cite{vijayakumar2018diverse}) }
BRR@10 measures whether the forgotten capability remains recoverable under beam search.
For each utterance $x \in \mathcal{D}_F$, we decode conditioned on the ground truth
intent prefix $I_F$. Let $y$ be the ground truth semantic frame and
$\hat{y}^{(k)}$ the $k$-th beam hypothesis with beam size $K=10$. We define:
\begingroup
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\abovedisplayshortskip}{1pt}
\setlength{\belowdisplayshortskip}{2pt}
\[
\mathrm{BRR@10}
=
\frac{1}{|\mathcal{D}_F|}
\sum_{(x,y)\in \mathcal{D}_F}
\mathbf{1}
\left(
\exists\, k \le K :
\hat{y}^{(k)} = y
\right)
\]
\endgroup
% The comparison between $\hat{y}^{(k)}$ and $y$ is performed using exact semantic frame matching.
\vspace{-0.1cm}
\noindent \textbf{Semantic Similarity \cite{reimers2019sentence} }To capture residual behavior beyond exact matching, we compute cosine similarity between embeddings $E(\hat{y})$ and $E(y)$, where $E(\cdot)$ maps structured semantic frames to vector representations. The cosine similarity is defined as follows:
\vspace{-0.1cm}
\begingroup
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{2pt}
\setlength{\abovedisplayshortskip}{1pt}
\setlength{\belowdisplayshortskip}{2pt}
\[
{\tiny
\mathrm{Sim}(\hat{y}, y)
=
\frac{E(\hat{y})^\top E(y)}
{\lVert E(\hat{y}) \rVert_2 \, \lVert E(y) \rVert_2}
}
\]
\endgroup
A higher $\mathrm{BRR}@10$ indicates recoverable capability under beam search, while higher semantic similarity reflects stronger alignment with the ground-truth frame despite surface differences.
% \vspace{-0.2cm}
\subsection{Baselines }
\vspace{-0.1cm}
As no prior work explicitly studies capability unlearning in end-to-end SLU, we adapt canonical machine unlearning methods for our setting. We evaluate the Gradient Ascent (GA) family, which maximizes the loss on the forget set to disrupt target representations \cite{golatkar2020eternal,yao2024large}, along with two stabilized variants: GA+GD, which additionally trains on retain data, and GA+KL, which applies KL regularization to the original model. We further include Negative Preference Optimization (NPO) and NPO+KL, which penalizes distribution alignment with forget-set outputs via a preference objective, and Random Label (RLabel), which replaces forget-set labels with random targets to simulate capability removal \cite{graves2021amnesiac}.
\begin{figure}[t!]
\centering
\begin{subfigure}[t]{0.46\columnwidth}
\centering
\includegraphics[height=0.78\columnwidth, trim=10pt 0pt 0pt 0pt, clip]{Figures/df_lambda_sensitivity_standard.png}
\vspace{-0.6cm}
\caption{Effect on Forget Set}
\label{fig:twoinrow:a}
\end{subfigure}\hfill
\begin{subfigure}[t]{0.46\columnwidth}
\centering
% trim = left bottom right top (crop LEFT side here)
\includegraphics[height=0.78\columnwidth, trim=39pt 0pt 0pt 0pt, clip]{Figures/dr_lambda_sensitivity_standard.png}
\vspace{-0.6cm}
\caption{Effect on Retain Set}
\label{fig:twoinrow:b}
\end{subfigure}
\vspace{-0.25cm}
\caption{\textbf{Effect of the binding regularizer $\lambda_{bind}$}. Increasing $\lambda_{bind}$ suppresses the target capability on the forget set ($D_F$), reducing all metrics, while performance on the retain set ($D_R$) remains largely stable.}
\label{fig:twoinrow}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.85\columnwidth]{Figures/beam_position_plot.png}
\vspace{-0.3cm}
\caption{\textbf{Exact recovery across beam positions.} We report the Beam Retrieval Rate (BRR@10), evaluating whether the correct semantic frame appears within the top-K beam hypotheses. Higher values indicate stronger target recoverability.}
\label{fig:beam_exact_recovery}
\end{figure}
\vspace{-0.35cm}
\section{Results and Analysis}
\subsection{Comparison with Baselines}
\vspace{-0.2cm}
% We evaluate whether slot content associated with the target intent can still be generated when the intent prefix is provided at test time, while the performance on non-target intents is preserved. As shown in Table~\ref{tab:02}, gradient and preference-based baselines (GA, GA+KL, NPO) substantially reduce marginal intent prediction on the forget set ($D_F$). However, their behavioral recovery scores remain high (BRR@10 $>$87.6, Sim$>$86.7), indicating that intent-conditioned slot content can still be recovered under forced-prefix decoding. In contrast, \textit{\textbf{BSU (ours)}} leads to pronounced reductions across all forget-set metrics. For NeMo Conformer-Transformer, BRR@10 decreases from 92.64 to 22.1 and semantic similarity from 90.14 to 24.8. Comparable reductions are observed with the SSL-initialized Conformer (BRR@10 from 85.42 to 16.3 and Sim from 89.27 to 18.7). The same trend is observed on SpeechMassive. At the same time, retain-set performance remains stable ($I_F=96.9$, $F1=92.6$), indicating that the removal effect is selective rather than degrading general performances.
We evaluate whether slot content associated with the target intent can still be generated when the intent prefix is provided at test time, while the performance on non-target intents is preserved. As shown in Table~\ref{tab:02}, gradient and preference-based baselines (GA, GA+KL, NPO) reduce marginal intent prediction on the forget set ($D_F$) in several settings. However, their forced-prefix recovery scores often remain high, indicating that intent-conditioned slot content can still be recovered under forced-prefix decoding. In contrast, \textit{\textbf{BSU (ours)}} leads to pronounced reductions across all forget-set metrics. For SLURP with NeMo Conformer-Transformer, BRR@10 decreases from 92.64 to 22.10 and semantic similarity decreases from 90.14 to 24.80. Comparable reductions are observed with the SSL-initialized Conformer, where BRR@10 decreases from 84.42 to 16.30 and semantic similarity decreases from 83.27 to 18.70. The same trend is observed on SpeechMassive, where BSU also reduces BRR@10 and semantic similarity across both model initializations. At the same time, retain-set performance remains stable relative to the original and retrain baselines in most settings, indicating selective unlearning.
% \vspace{-0.2cm}
\subsection{Residual Capability Analysis}
% \vspace{-0.1cm}
\noindent \textbf{Random Space Analysis. }We include a Random Space (RS) ablation in Table~\ref{tab:02} to test whether unstructured perturbations in the representation space can induce forgetting. RS perturbs representations along random directions that are not aligned with the subspace encoding the target intent–slot dependency. Consequently, the perturbation does not systematically disrupt the underlying capability, resulting in limited suppression.
\vspace{0.3em}
\noindent \textbf{Binding Regularizer Sensitivity. }Figure~\ref{fig:twoinrow} analyzes as $\lambda_{bind}$ increases, performance on the forget split ($I_F$, $F1_F$, BRR@10) decreases steadily, indicating progressively stronger suppression of the target capability. In contrast, metrics on the retain split remain largely stable across the same range of $\lambda_{bind}$.
\vspace{0.3em}
\noindent \textbf{Residual Capability Recovery. }Figure~\ref{fig:beam_exact_recovery} demonstrates whether the target capability can be regenerated during decoding by leveraging the beam position of the first exact semantic-frame recovery. The stacked bars show the fraction of forget-set ($D_F$) samples whose correct frame first appears at beam position $k$ ($1 \leq k \leq 10$). The original model recovers the correct frame for a large portion of samples at early beam positions, indicating strong memorization of the target behavior. Both retraining and BSU significantly lower recovery across all beam positions, indicating that exact recovery is strongly reduced, even along alternative decoding paths.
\vspace{-0.2cm}
\section{Conclusion}
\vspace{-0.2cm}
In this work, we show that suppressing marginal intent prediction alone does not eliminate the conditional mapping governing slot generation, leading to capability persistence under forced-prefix decoding. We propose \textit{Binding Subspace Unlearning (BSU)}, which removes this dependency by targeting representation-level binding directions. Experiments show substantial reductions in recoverable slot behavior while preserving retained-intent performance. Future work includes extending this approach to broader generative tasks and developing principled capability-level unlearning methods for privacy-centric applications.
\section{Acknowledgments}
We acknowledge the institutional and computational support provided by the Department of Data Science and Engineering, Indian Institute of Science Education and Research Bhopal.
\section{Generative AI Use Disclosure}
Generative AI tools were used only for language editing and polishing. All scientific content, experimental design, analyses, results, and conclusions were developed and verified by the authors, who take full responsibility for the content of this paper.
% \section{Acknowledgments}
% {\color{blue}Acknowledgments should be included only in the camera-ready version, not in the version submitted for review. For regular papers, pages 5 and 6, and for long papers, pages 9 and 10, are reserved exclusively for acknowledgments, disclosures of the use of generative AI tools, and references. No other content may appear on these pages. Any appendices must be contained within the first four pages for regular papers and within the first eight pages for long papers.
% Acknowledgments and references may begin on an earlier page if space permits.}
% \ifcameraready
% The Interspeech 2026 organizers
% \else
% The authors
% \fi
% would like to thank ISCA and the organizing committees of past Interspeech conferences for their help and for kindly providing the previous version of this template.
% {\color{blue}
% \section{Generative AI Use Disclosure}
% The extent of Generative AI use must be disclosed. This section may be in the 5th or 6th pages of regular papers, or the 9th or 10th pages of long papers. ISCA policy says: \textit{All (co-)authors must be responsible and accountable for the work and content of the paper, and they must consent to its submission. Any generative AI tools cannot be a co-author of the paper. They can be used for editing and polishing manuscripts, but should not be used for producing a significant part of the manuscript}.}
% \section{Reference Format}
% {\color{blue}It is ISCA policy that papers submitted to Interspeech should refer to peer-reviewed publications. References to non-peer-reviewed publications (including public repositories such as arXiv, Preprints, and HAL, software, and personal communications) should only be made if there is no peer-reviewed publication available, and should be kept to a minimum.
% References should be in standard IEEE format (follows the IEEEtran format), numbered in order of appearance, for example \cite{Davis80-COP} is cited before \cite{Rabiner89-ATO}. For longer works such as books, provide a single entry for the complete work in the References, then cite specific pages \cite[pp.\ 417--422]{Hastie09-TEO} or a chapter \cite[Chapter 2]{Hastie09-TEO}. Multiple references may be cited in a list \cite{Smith22-XXX, Jones22-XXX}.}
\bibliographystyle{IEEEtran}
\bibliography{mybib}
\end{document}
%%% Local Variables:
%%% mode: LaTeX
%%% TeX-master: t
%%% End: