Tag
This paper provides causal evidence that large language models acquire negative linguistic knowledge (what not to say) through statistical preemption, a mechanism from Construction Grammar, by showing that manipulating competing-form frequencies via fine-tuning shifts preemption behavior in predicted directions.