Instructions shape Production of Language, not Processing

This research paper explores a production-centered mechanism in language models, revealing an asymmetry between language processing and production. It shows that instructions significantly shape information in output tokens, but not in sample tokens, correlating strongly with model behavior.