\Needspace{17\baselineskip}
\subsection{Vision retained at constant memory}
\label{sec:vision}

EdgeRock-4B extends its constant-memory design to text-and-image reasoning. Across six visual benchmarks covering perception, chart and diagram understanding, and visual question answering, it reaches an \textbf{84.9\%} equal-weight mean answer-level accuracy, compared with 89.1\% for Qwen3.5-4B, retaining 95.3\% (Table~\ref{tab:visual-results}).

\begin{table}[!htbp]
\centering
\caption{Visual answer-level accuracy (\%), including MME. Retained gives EdgeRock-4B as a percentage of Qwen3.5-4B. Appendix~\ref{app:protocol} gives standard MME points.}
\label{tab:visual-results}
\begin{tabularx}{\linewidth}{@{}Yrrr@{}}
\toprule
Benchmark & EdgeRock-4B & Qwen3.5-4B & Retained (\%)\\
\midrule
MMBench EN~\cite{mmbench} & 88.3 & 90.1 & 98.0 \\
MMBench CN~\cite{mmbench} & 89.0 & 91.4 & 97.4 \\
POPE~\cite{pope} & 85.4 & 88.8 & 96.2 \\
MME~\cite{mme} & 83.0 & 86.8 & 95.6 \\
ChartQA~\cite{chartqa} & 83.9 & 90.3 & 92.9 \\
AI2D~\cite{ai2d} & 80.0 & 87.0 & 92.0 \\
\midrule
Equal-weight mean & 84.9 & 89.1 & 95.3 \\
\bottomrule
\end{tabularx}
\end{table}