Sparks of Pure Competence in LLMs: The Case of Syntactic Center Embedding in English

Linguistic theory distinguishes between competence and performance: the competence grammar ascribed to humans is not always clearly observable, because of performance limitations. This raises the possibility that an LLM, if it is not subject to the same performance limitations as humans, might exhibit behavior closer to a pure instantiation of the human competence model. We explore this in the case of syntactic center embedding, where, the competence grammar allows unbounded center embedding, although humans have great difficulty with any level above one. We study this in four LLMs, and we find that the most powerful model, GPT-4, does appear to be approaching pure competence, achieving high accuracy even with 3 or 4 levels of embeddings, in sharp contrast to humans and other LLMs.