Examining Compositional Behavior in Large Language Models

A research project supervised by Professor Vivek Srikumar on how large language models solve compositional tasks. The project utilized the HuggingFace library and Slurm scripts for experiments with Flan-T5 and Llama 2-Chat models. After various experiments involving multiplication and word list tasks, I determined that prompting with subtask examples did not reliably improve performance on compositional tasks.