Presentation description

Transformer-based large language models (LLMs) have seen prolific usage in natural language processing (NLP) tasks, search engines, and code generation. Despite their ability to compose intricate responses, LLMs exhibit reasoning-based limitations in compositional tasks. These are tasks that are complex enough that they require multiple subtasks to complete them. We explore the idea of whether prompting an LLM with in-context subtask examples can improve compositionality.