summaryrefslogtreecommitdiff
path: root/paper
diff options
context:
space:
mode:
Diffstat (limited to 'paper')
-rw-r--r--paper/figures/fig4_penalty_rescue.pdfbin32527 -> 34190 bytes
-rw-r--r--paper/figures/fig5_cross_arch_summary.pdfbin31765 -> 32029 bytes
-rw-r--r--paper/figures/render_fig4_penalty_rescue.py2
-rw-r--r--paper/figures/render_fig5_cross_arch.py6
-rw-r--r--paper/main.pdfbin484918 -> 486620 bytes
-rw-r--r--paper/main.tex4
6 files changed, 6 insertions, 6 deletions
diff --git a/paper/figures/fig4_penalty_rescue.pdf b/paper/figures/fig4_penalty_rescue.pdf
index 0dcdce7..e703f08 100644
--- a/paper/figures/fig4_penalty_rescue.pdf
+++ b/paper/figures/fig4_penalty_rescue.pdf
Binary files differ
diff --git a/paper/figures/fig5_cross_arch_summary.pdf b/paper/figures/fig5_cross_arch_summary.pdf
index 93c0676..00267a7 100644
--- a/paper/figures/fig5_cross_arch_summary.pdf
+++ b/paper/figures/fig5_cross_arch_summary.pdf
Binary files differ
diff --git a/paper/figures/render_fig4_penalty_rescue.py b/paper/figures/render_fig4_penalty_rescue.py
index b7089ec..ad21a12 100644
--- a/paper/figures/render_fig4_penalty_rescue.py
+++ b/paper/figures/render_fig4_penalty_rescue.py
@@ -36,7 +36,7 @@ no_pen = [0.609, 0.308]
with_pen = [0.530, 0.363]
shallow = 0.349
-fig, axes = plt.subplots(1, 3, figsize=(13, 3.5))
+fig, axes = plt.subplots(1, 3, figsize=(13, 6.0))
# Panel A: trajectory
ax = axes[0]
diff --git a/paper/figures/render_fig5_cross_arch.py b/paper/figures/render_fig5_cross_arch.py
index e163956..9ad9ce2 100644
--- a/paper/figures/render_fig5_cross_arch.py
+++ b/paper/figures/render_fig5_cross_arch.py
@@ -26,7 +26,7 @@ dfa = np.array([
bp = np.zeros_like(dfa) # BP: passes everywhere
-fig, axes = plt.subplots(1, 2, figsize=(11, 3.5))
+fig, axes = plt.subplots(1, 2, figsize=(11, 4.2))
for ax, mat, title in [(axes[0], bp, "BP-trained: protocol passes"),
(axes[1], dfa, "DFA-trained: protocol verdict by architecture")]:
@@ -42,8 +42,8 @@ for ax, mat, title in [(axes[0], bp, "BP-trained: protocol passes"),
ax.set_yticklabels(arches, fontsize=9)
ax.set_title(title, fontsize=10)
-# Highlight the key finding
-axes[1].text(0.5, -1.0, "Key finding: diagnostic (b) BP-grad-floor fires only on terminal-LN architectures.\n"
+# Highlight the key finding — place well below the multiline x/y tick labels to avoid overlap
+axes[1].text(0.5, -1.55, "Key finding: diagnostic (b) BP-grad-floor fires only on terminal-LN architectures.\n"
"Across the 5 architecture cases tested, (b) is restricted to the with-terminal-LN family.",
ha="center", fontsize=9, style="italic", transform=axes[1].transAxes)
diff --git a/paper/main.pdf b/paper/main.pdf
index 70915bb..7d5dfea 100644
--- a/paper/main.pdf
+++ b/paper/main.pdf
Binary files differ
diff --git a/paper/main.tex b/paper/main.tex
index fa9c12b..8486315 100644
--- a/paper/main.tex
+++ b/paper/main.tex
@@ -133,14 +133,14 @@ The architecture comparison sharpens the scope of the critique. In the terminal-
\begin{figure}[t]
\centering
-\includegraphics[width=0.92\linewidth]{figures/fig4_penalty_rescue.pdf}
+\includegraphics[width=\linewidth]{figures/fig4_penalty_rescue.pdf}
\caption{Penalty intervention view of the two modes: penalization rescues residual-stream scale and restores a measurable but still partial deep-layer credit signal, clarifying that numerical rescue and credit-quality rescue are related but distinct.}
\label{fig:penalty_rescue}
\end{figure}
\begin{figure}[t]
\centering
-\includegraphics[width=0.92\linewidth]{figures/fig5_cross_arch_summary.pdf}
+\includegraphics[width=\linewidth]{figures/fig5_cross_arch_summary.pdf}
\caption{Cross-architecture summary over ResMLP, ViT-Mini, StudentNet, and CNN: activation-growth failures recur across architectures, while gradient-floor failures appear in the terminal-normalized settings audited here.}
\label{fig:cross_arch_summary}
\end{figure}