Update index.html

Author: MJXWANG

index.html

@@ -483,6 +483,36 @@ <h3 class="title is-4 has-text-centered">Result Summary</h3>
   </div>
 </section>
 
+          
+<section id="implications" class="section">
+  <div class="container is-max-desktop">
+    <h2 class="title is-3 has-text-centered">🧭 Implications & Future Work</h2>
+
+    <div class="content has-text-justified" style="max-width: 880px; margin: 0 auto;">
+      
+      <h3 class="title is-5">🧠 Implications for Cognitive AI</h3>
+      <ul>
+        <li><strong>Emergent Congruency:</strong> The classic Stroop and Flanker effects emerge even in untrained models, suggesting interference control arises from foundational pretraining dynamics.</li>
+        <li><strong>Limits of Scaling:</strong> Even 110B-parameter models fail on deeply nested conflicts, pointing to architectural or training regime bottlenecks beyond sheer size.</li>
+        <li><strong>Shared Control Mechanisms:</strong> Strong correlations between letter- and number-Flanker scores (r = 0.96) support the idea of a unified control construct within VLMs.</li>
+      </ul>
+
+      <h3 class="title is-5">🔍 Open Questions</h3>
+      <ul>
+        <li>How do specific aspects of pretraining data influence the emergence of conflict control?</li>
+        <li>What inductive biases or training interventions are needed to overcome hierarchical interference?</li>
+        <li>Can VLMs develop temporally extended control structures analogous to executive functions?</li>
+      </ul>
+
+      <h3 class="title is-5">🧪 Methodological Contribution</h3>
+      <p>
+        This study introduces a psychophysics-inspired paradigm for evaluating control behavior in VLMs. 
+        By using tightly controlled stimuli and separating standard vs. hierarchical conflict, we provide 
+        mechanistic insights and open a new path toward studying cognitive control in AI systems.
+      </p>
+    </div>
+  </div>
+</section>