FURI | Fall 2023

Diversity Measures: Domain-Independent Proxies for Failure in Language Model Queries

Data icon, disabled. Four grey bars arranged like a vertical bar chart.

This research examines the correlation between a language model’s uncertainty — quantified via “diversity measures” — and its reasoning abilities. The research team evaluates aspects of the model’s diversity measures that are strong indicators of it being correct. Doing this provides valuable insights into the capabilities of language models and allows users to further improve the user’s prompts to enhance model performance.

Student researcher

Noel Ngu

Computer science

Hometown: Kuching, Sarawak, Malaysia

Graduation date: Fall 2023