Tag
This paper evaluates the consistency and specificity of language model circuits, finding that while circuits are consistent within tasks, they lack task-specificity due to substantial overlap across different tasks.