Language Models and Generative Linguistic Theories Are More Compatible Than They Appear
This academic paper, submitted to arXiv by R. Thomas McCoy on May 11, 2026, challenges the prevailing view that neural language models (LMs) exclusively support gradient, usage-based linguistic theories. While previous research by Futrell and Mahowald (2025) framed LM success as evidence for usage-based accounts, McCoy argues that LMs can also instantiate theories based on formal structures, characteristic of the generative linguistic tradition. This argument significantly expands the theoretical space available for testing with language models. By demonstrating that LMs are not limited to supporting one linguistic paradigm, the study suggests a potential pathway for reconciling the long-standing divide between usage-based and generative accounts in linguistics. The work contributes to the fields of computational linguistics and artificial intelligence, offering new perspectives on how modern AI tools can be utilized to validate and explore diverse theoretical frameworks in human language understanding. It highlights the versatility of neural networks in modeling complex linguistic structures beyond simple statistical patterns.
Wire timeline
Language Models and Generative Linguistic Theories Are More Compatible Than They Appear
This academic paper, submitted to arXiv by R. Thomas McCoy on May 11, 2026, challenges the prevailing view that neural language models (LMs) exclusively support gradient, usage-based linguistic theories. While previous research by Futrell and Mahowald (2025) framed LM success as evidence for usage-based accounts, McCoy argues that LMs can also instantiate theories based on formal structures, characteristic of the generative linguistic tradition. This argument significantly expands the theoretical space available for testing with language models. By demonstrating that LMs are not limited to supporting one linguistic paradigm, the study suggests a potential pathway for reconciling the long-standing divide between usage-based and generative accounts in linguistics. The work contributes to the fields of computational linguistics and artificial intelligence, offering new perspectives on how modern AI tools can be utilized to validate and explore diverse theoretical frameworks in human language understanding. It highlights the versatility of neural networks in modeling complex linguistic structures beyond simple statistical patterns.
cs.AI updates on arXiv.org