Bias Amplification in Language Model Evolution: An Iterated Learning Perspective