A 26%-54% mismatch rate indicates that the model knows what it should do but cannot do it; the gap between cognition and action is deeper than imagined.

This interpretability study focuses on proxy tools, revealing that although the model can recognize when to call a tool, actual calls often fail, with a mismatch rate of 26%–54%. The issue centers on the transformation from cognition to action, not cognition itself. Internal signals can be decoded, but the final token mechanism in the later layers causes signal rotation, making it almost orthogonal to the action. The research aims to predict intervention effects, suggesting that attributions such as insufficient prompting or training may overlook the geometric structure of the later layers, thereby explaining the performance ceiling observed in tool usage A/B testing.