Update index.html

Author: MJXWANG

index.html

@@ -302,10 +302,10 @@ <h1 class="title is-1">🧠 Machine Psychophysics: Cognitive Control in Vision
       <div class="column is-four-fifths">
         <h2 class="title is-3">Abstract</h2>
         <div class="content has-text-justified">
-          <p><strong>Cognitive control</strong> refers to the ability to flexibly coordinate thought and action. Conflict-task paradigms benchmark this faculty by contrasting congruent and incongruent trials.</p>
-          <p>We conduct the first <strong>large-scale psychophysics evaluation</strong> of Vision–Language Models, testing <strong>108 models</strong> on Stroop, Letter- and Number-Flanker tasks plus their more demanding “<em>Squared</em>” variants — <strong>across 2 220 trials, plus 238 control trials</strong>.</p>
-          <p>Models show human-like congruency patterns, yet collapse when conflicts are hierarchical. Accuracy grows <em>roughly log-linearly</em> with parameter count, echoing resource-limited curves in human forced-response studies.</p>
-          <p>The benchmark demonstrates that executive-function signatures can emerge from general-purpose learning at scale while revealing clear gaps for future research.</p>
+          <p><strong>Cognitive control</strong> refers to the ability to flexibly coordinate thought and action in pursuit of internal goals. Conflict-task paradigms benchmark this faculty by contrasting congruent and incongruent trials.</p>
+          <p>We evaluate <strong>108 vision–language models</strong> on Stroop, Letter- and Number-Flanker tasks and their more demanding “<em>Squared</em>” variants — <strong>across 2,220 structured trials and 238 control trials</strong>. Models reproduce human-like congruency effects and, critically, show <strong>robust inter-model variation</strong> that reflects differential sensitivity to interference.</p>
+          <p>Letter- and Number-Flanker scores are <strong>highly correlated</strong> (r = 0.96), indicating stable, convergent traits of control. Furthermore, accuracy improves <em>log-linearly</em> with parameter scale, aligning with human forced-response processing curves.</p>
+          <p>These results support the emergence of control mechanisms from general-purpose associative learning and introduce a framework for measuring trait-like cognitive properties in large-scale AI systems.</p>
         </div>
       </div>
     </div>