Tag
This paper examines how program inspection of Transformer Programs reveals failures hidden by token accuracy metrics across nine algorithmic tasks, showing how inspecting extracted Python code can guide training data improvements and boost sequence accuracy.