Author ORCID Identifier

https://orcid.org/0000-0001-6547-9137

Date of Award

Summer 6-14-2026

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Soroush Vosoughi

Second Advisor

Yaoqing Yang

Third Advisor

Adam Breuer

Abstract

Large language models (LLMs) are increasingly deployed in high-stakes scenarios, yet their tendency to encode and replicate human-like social biases threatens their safe and fair application in these settings. Current approaches for measuring these biases in LLMs often rely on large, manually annotated datasets; however, these methods often consider a limited set of languages and cultural contexts and are operationalized in ways that limit the granularity and robustness of reported metrics. This thesis aims to expand the scope of bias benchmarking along three dimensions. Firstly, we develop a human-LLM collaborative annotation framework to more efficiently construct stereotype examination datasets for low-resource languages and cultures. We apply this framework to construct EspanStereo, a stereotype examination dataset of 2,690 instances specific to five countries in the Spanish-speaking world. Secondly, we show that preprocessing-based debiasing techniques can redistribute, rather than eliminate, encoded biases in LLMs in ways that are frequently invisible to current benchmarks. Finally, we develop a semantics-based debiasing method with an associated gold-standard benchmark that offers a more granular means of stereotype examination than fixed-sequence benchmarks.

Available for download on Saturday, June 05, 2027

Share

COinS