LLMs Are Bad at Being Forced

The further you push language models from their natural output patterns, the worse they perform. Understanding why reveals fundamental truths about how these systems actually work.