Tag
The paper introduces a bias depth score to distinguish stable 'deep biases' from prompt-dependent 'shallow biases' in large language models, showing that deep biases are more persistent and harder to remove.