% ============================================================
% Placement: Appendix (Experimental Setup reproducibility)
% ============================================================
\begin{table}[t]
\centering
\small
\begin{adjustbox}{max width=\columnwidth}
\begin{tabular}{llccc}
\toprule
\textbf{Model} & \textbf{Access} & \textbf{Temp.} & \textbf{Top-$p$} & \textbf{Max Tokens} \\
\midrule
GPT-5.2 & API & 0.0 & 1.0 & 4,096 \\
GPT-5.4-mini & API & 0.0 & 1.0 & 4,096 \\
\midrule
DeepSeek-v3.2 & API & 0.0 & 1.0 & 4,096 \\
Qwen2.5-Coder-32B-Instruct & local (vLLM) & 0.0 & 1.0 & 4,096 \\
Qwen2.5-Coder-7B-Instruct & local (vLLM) & 0.0 & 1.0 & 4,096 \\
Qwen2.5-Coder-3B-Instruct & local (vLLM) & 0.0 & 1.0 & 4,096 \\
Qwen2.5-Coder-0.5B-Instruct & local (vLLM) & 0.0 & 1.0 & 4,096 \\
Llama-3.2-3B-Instruct & local (vLLM) & 0.0 & 1.0 & 4,096 \\
Llama-3.2-1B-Instruct & local (vLLM) & 0.0 & 1.0 & 4,096 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\caption{Decoding configurations for all evaluated models. All models use greedy decoding. Local models are served via vLLM.}
\label{app:model_decoding_config}
\end{table}