关键词: 人工智能, 大语言模型, 语言能力, 语法能力, 语言习得, 句法加工

Abstract: After reviewing researches on the development of linguistic competence of Large Language Models（LLMs）and comparing the different characteristics between LLMs and human speech learning，this study explores the evaluation of LLMs’linguistic competence and its theoretical implications from multiple dimensions，including the learning environment and mechanism，the measurement of languagespecific generalization ability，and the assessment of grammatical competence. It can be found that：In terms of learning environment，LLMs achieve efficient statistical generalization with massive single-modal text input，while humans develop language capacity in multi-modal interactions with higher ecological validity； their differences are complementary. Regarding the core assumption of genetic theory of language，the results of ablation experiment，unsupervised and supervised tests reveal that although LLMs lack the prior grammatical specificity of humans，they can reproduce some grammatical rules through statistical models. The assessment of grammatical competence indicates that although LLMs can acquire surface syntactic structures，there are still significant limitations in modeling human-specific features such as deep recursion and semantic-syntactic interfaces. Meanwhile，the emergent ability of LLMs poses a dual challenge to the theory of stimulus scarcity and the genetic theory of language；it promotes the paradigm fusion of computational linguistics with theoretical linguistics，cognitive science and other fields. In the future， the assessment of LLMs’ language capabilities needs to focus on the cognitive decoupling mechanism between language forms and functions，so as to explore the collaborative approaches of interdisciplinary methodologies and clarify LLMs’language capability boundaries.

Key words: artificial intelligence （AI）, Large Language Model （LLM）, linguistic competence, grammatical competence, language acquisition, syntactic processing