LLMs highlight the importance of interaction in human language learning

Abstract Recent years have seen large language models (LLMs) achieve impressive performance with core linguistic abilities. This can be taken as a demonstration that, contrary to long-held assumptions about innate linguistic constraints, language can be learned through statistical learning from linguistic input alone. However, statistical language learning in these models differs from human learners in a crucial way: human language acquisition evolved not simply as passive absorption of linguistic input but is instead fundamentally interactive, guided by continuous feedback and social cues. Recent advances in LLM engineering have introduced an additional step in model training that utilizes more human-like feedback in what has come to be known as reinforcement learning from human feedback (RLHF). This procedure results in models which more closely mirror human linguistic behaviors – even reproducing characteristic human-like errors. We argue that the way RLHF changes the behavior of LLMs highlights how communicative interaction and socially informed feedback, in addition to input-driven statistical learning, can explain fundamental aspects of language learning. In particular, we take LLMs as models of “idealized statistical language learners” and RLHF as a form of “idealized language feedback”, arguing that this perspective offers valuable insights into our understanding of human language development.