docs: update novelty section

Author: 404Simon

presentation_final/sections/05_novelty.tex

@@ -5,17 +5,18 @@
       hysteresis-based pausing, validated against ground-truth energy models
       and cross-model scaling expectations.
 
-    \vspace{0.2cm}
-    \item \textbf{Adaptive grid-search optimizer} that efficiently explores the
-      $(\theta_p, \theta_r, t_s)$ policy space, converging within
-      $\leq$10 iterations.
+      \vspace{0.2cm}
+    \item \textbf{Adaptive grid-search optimizer} that efficiently
+      explores the $(\theta_p, \theta_r, t_s)$ policy space,
+      converging within $\leq$6-7 iterations. (a discrete, 3D
+      variation of the Golden Section Search)
 
-    \vspace{0.2cm}
+      \vspace{0.2cm}
     \item \textbf{Comprehensive empirical study} covering 2~SOTA models
       (DeepSeek~V3, Kimi~K2) $\times$ 5~grid regions of replay data
       --- the first systematic Pareto frontiers for carbon-aware LLM training.
 
-    \vspace{0.2cm}
+      \vspace{0.2cm}
     \item \textbf{Actionable design rule:} the optimal hysteresis policy uses a
       narrow margin ($\theta_p - \theta_r < 16$~gCO$_2$eq/kWh).
       Wide hysteresis is never Pareto-optimal.