| \begin{table}[t] |
| \centering\small |
| \caption{Diagnosis-enabled interventions on the primary cell (\texttt{qwen25-7b} / GPTQ w4 / MATH-500). \emph{Prompt-prefix $k$} prepends the first $k$ FP16 reference steps to the quantized model's prompt (no training). The \emph{unrestricted} sweep injects every FP16 prefix; the \emph{FP16-correct-only} sweep injects a prefix only when the source FP16 trace reached the correct final answer, removing partial-solution leakage. \emph{QLoRA restored} is the adapter from \S\ref{sec:results-restoration}. Accuracy in \%, 95\% bootstrap CIs in brackets.} |
| \label{tab:interventions} |
| \begin{tabular}{@{}lcc@{}} |
| \toprule |
| Intervention & Acc (\%) & 95\% CI \\ |
| \midrule |
| Quantized baseline & 69.4 & -- \\ |
| \midrule |
| \multicolumn{3}{l}{\textit{Unrestricted prefix (diagnostic)}} \\ |
| Prompt-prefix $k=0$ & 70.2 & [66.2, 74.0] \\ |
| Prompt-prefix $k=1$ & 75.2 & [71.4, 78.8] \\ |
| Prompt-prefix $k=2$ & 80.6 & [77.0, 84.0] \\ |
| Prompt-prefix $k=3$ & 87.8 & [85.0, 90.6] \\ |
| Prompt-prefix $k=4$ & 91.4 & [89.0, 93.8] \\ |
| \midrule |
| \multicolumn{3}{l}{\textit{FP16-correct-only prefix (honest)}} \\ |
| Prompt-prefix $k=0$ (FP16-correct) & 70.4 & [66.4, 74.2] \\ |
| Prompt-prefix $k=1$ (FP16-correct) & 73.6 & [69.6, 77.4] \\ |
| Prompt-prefix $k=2$ (FP16-correct) & 75.6 & [71.8, 79.2] \\ |
| Prompt-prefix $k=3$ (FP16-correct) & 77.6 & [74.0, 81.2] \\ |
| Prompt-prefix $k=4$ (FP16-correct) & 78.0 & [74.2, 81.4] \\ |
| \midrule |
| \textbf{QLoRA restored (ours)} & 74.8 & -- \\ |
| FP16 reference & 72.8 & -- \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
