BLOOM Large Language Models and the Chomsky Hierarchy

Sumit Kumar Jha, Rickard Ewetz, Alvaro Velasquez, Laura Pullum, Susmit Jha

Keywords: Large Language Models, Chomsky Hierarchy

Abstract: We study the performance of BLOOM large language models of different sizes on understanding 12 different languages from the Chomsky hierarchy using few-shot prompts. We investigate whether an increase in the complexity of the languages learned by the larger models can be characterized using the Chomsky hierarchy. We first show that prompting in BLOOM models enables reasoning with a good accuracy on language tasks as diverse as stack manipulation, string reversal, odds first, and interlocked pairing, when the queries are over short strings, that is, small bitwidth bit-vectors from the language. Second, we discover that the two largest models have the highest accuracy on such tasks for prompts with a fixed length, but smaller models are able to achieve similar accuracies with longer prompts. Unlike classical automata or grammar based approaches where algorithms for more complex languages in the Chomsky hierarchy can also recognize simpler languages, we find that the performance of the BLOOM large language models cannot be explained by the complexity of the languages in the Chomsky hierarchy.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)

TL;DR: The performance of the BLOOM large language models cannot be explained by the complexity of the languages in the Chomsky hierarchy.