5/ One result stood out. In one test the model correctly named the experiment and explained the Stroop task in detail. Then it failed the task anyway. Knowing the rule did not help it follow the rule.

→ View original post on X — @aihighlight