LLMs Fail Abstract Syntactic Generalization Compared to Humans

A study comparing linguistic productivity shows that LLMs require semantic associations to generalize syntactic forms, such as moving novel verbs from active to passive voice. Humans generalize based on abstract form regardless of meaning, whereas LLM success is tied to existing lexical distributions.