\begin{table}[t]
\caption{Controlled quantitative snapshot under 40\% instance-dependent noise (IDN). Test accuracy (\%, mean $\pm$ standard deviation) is transcribed from Table~1 of DISC \citep{li2023disc} for its CIFAR protocol. Asterisks mark results reimplemented by the DISC authors; the remaining baseline results were taken by that paper from its cited comparison source. This table documents one protocol and is not a universal method ranking.}\label{tab:controlled_quantitative}
\small
\setlength{\tabcolsep}{4pt}
\begin{tabular}{@{}L{0.31\textwidth}L{0.31\textwidth}C{0.14\textwidth}C{0.14\textwidth}@{}}
\toprule
Method & Principal mechanism & CIFAR-10 & CIFAR-100 \\
\midrule
CE$^{*}$ & Standard cross-entropy baseline & $67.64 \pm 0.26$ & $43.17 \pm 0.15$ \\
Forward T \citep{patrini2017loss} & Transition-matrix loss correction & $79.37 \pm 2.72$ & $42.80 \pm 1.01$ \\
GCE$^{*}$ \citep{zhang2018generalized} & Robust generalized cross entropy & $78.95 \pm 0.15$ & $45.69 \pm 0.14$ \\
Co-teaching \citep{han2018coteaching} & Peer small-loss sample exchange & $73.00 \pm 1.24$ & $23.21 \pm 0.57$ \\
DivideMix \citep{li2020dividemix} & Loss-mixture separation and semi-supervised reuse & $95.07 \pm 0.11$ & $76.08 \pm 0.35$ \\
DISC \citep{li2023disc} & Dynamic instance-specific selection and correction & $95.94 \pm 0.04$ & $78.44 \pm 0.19$ \\
\botrule
\end{tabular}
\end{table}