Tag
This paper investigates how training alignment objectives reshape linguistic features in large language models, finding that instruction-tuned systems collapse language entropy significantly more than scale would suggest, and that entropy regularization can mitigate this collapse.