\label{sec:delta}

\begin{figure}[!htbp]
\centering
\includegraphics[width=\linewidth]{backbone_delta.pdf}
\caption{\textbf{Capability after conversion.} Bars report EdgeRock-4B minus its Qwen3.5-4B source backbone in percentage points, ordered from largest to smallest. Blue highlights the eight benchmark gains; aligned columns give absolute scores (\%).}
\label{fig:delta}
\end{figure}

\subsection{In the range of compact models with growing memory}
\label{sec:peer-capability}
Against compact models whose histories grow, EdgeRock-4B reasons in the same range. On AIME 2024--2026, its 82.9\% lies between Granite4.2-3B (82.2\%) and Qwen3.5-4B (86.1\%), with MiniCPM5-2B at 87.8\%, and well above Nemotron-3-Nano-4B (57.1\%), LFM2.5-2.6B (56.2\%), Gemma4-E4B (40.1\%), and Phi-4-mini-reasoning (38.9\%) (Table~\ref{tab:peer-capability}; Appendix~\ref{app:growing-results}). EdgeRock-4B's advantage over MiniCPM5-2B is larger on MMLU-Pro and SuperGPQA, which test knowledge-intensive reasoning (Table~\ref{tab:peer-capability}).

EdgeRock-4B also leads LFM2.5-2.6B on \textbf{28 of 29 benchmarks} (Appendix~\ref{app:growing-results}).

\PeerCapabilityResults

\input{chapters/vision.tex}
\FloatBarrier