\begin{table*}[t]
\centering
\scriptsize
\setlength{\tabcolsep}{2.5pt}
\begin{tabular}{@{}>{\raggedright\arraybackslash}p{3.0cm}lcrcrrrr@{}}
\toprule
LLaDA-V pool & bench & $T{=}4$ $\rightarrow$ $+$ADR & ADR gain & $T{=}8{+}$ADR / $T{=}32$ & fragility $\Delta$ & early read $\Delta$ & own text & blank img. \\
\midrule
original pool (60 questions) & \pope{} & 0.660 $\rightarrow$ 0.783 & +0.123 {\scriptsize[+0.077,+0.170]} & 0.837 / 0.847 & +0.33 {\scriptsize[+0.25,+0.42]} & +0.25 {\scriptsize[+0.14,+0.36]} & 99.8\% & +0.000 \\
adversarial, 300 images & \pope{} & 0.707 $\rightarrow$ 0.793 & +0.087 {\scriptsize[+0.043,+0.130]} & 0.833 / 0.833 & +0.28 {\scriptsize[+0.20,+0.36]} & +0.24 {\scriptsize[+0.14,+0.34]} & 99.8\% & +0.001 \\
random, 300 images & \pope{} & 0.780 $\rightarrow$ 0.907 & +0.127 {\scriptsize[+0.083,+0.170]} & 0.913 / 0.907 & +0.20 {\scriptsize[+0.10,+0.30]} & +0.17 {\scriptsize[+0.03,+0.29]} & 99.7\% & +0.001 \\
adversarial $+$ noise ($\sigma{=}0.2$) & \pope{} & 0.580 $\rightarrow$ 0.720 & +0.140 {\scriptsize[+0.087,+0.193]} & 0.747 / 0.757 & +0.20 {\scriptsize[+0.13,+0.28]} & +0.08 {\scriptsize[$-$0.02,+0.17]} & 99.7\% & +0.001 \\
two-hop (original pool) & \gqa{} & 0.390 $\rightarrow$ 0.410 & +0.020 {\scriptsize[$-$0.027,+0.067]} & 0.410 / 0.423 & +0.22 {\scriptsize[+0.15,+0.29]} & +0.35 {\scriptsize[+0.26,+0.44]} & 93.9\% & +0.010 \\
\bottomrule
\end{tabular}
\caption{\textbf{The LLaDA-V results on fresh pools and under noise} ($n{=}300$ each; the fresh \pope{} pools use one question per image, $300$ distinct images; the original pool repeats 60 questions). The deferred round, schedule fragility and the substitution result replicate on every \pope{} pool, including under Gaussian noise that lowers accuracy by $7.7$ points; under that noise the early read's gain shrinks to $+0.08$ (not significant), and on \gqa{} the deferred round's accuracy gain is not significant, because the schedule loss there is small. Supp.~\ref{app:ext_gen} gives every arm with intervals.}
\label{tab:robust}
\end{table*}