Shreya Havaldar
University of Pennsylvania
shreyah@seas.upenn.edu
Bio
Shreya Havaldar is a Ph.D. candidate in Computer Science at the University of Pennsylvania, advised by Lyle Ungar and Eric Wong. Her research, supported by the NSF Graduate Research Fellowship, focuses on human‑centered NLP, particularly the evaluation and mitigation of cultural bias in large language models (LLMs). She earned B.S. degrees in Computer Science and Applied and Computational Mathematics from the University of Southern California, where she worked with Morteza Dehghani. She has previously interned at Google DeepMind and Spotify Research and leads the Women in Machine Learning (WiML) organization at Penn.
Areas of Research
- Natural Language and Speech Processing
Towards Evaluating and Mitigating Cultural Bias in LLMs
Large Language Models (LLMs) play a significant role in human communication, yet they predominantly reflect Anglocentric norms and beliefs. This bias can lead to culturally insensitive outputs and erode trust among non‑Western users. My research addresses this challenge by making LLMs more reliable across cultures and communities. First, I develop methods to explain how LLMs represent culture: by grounding evaluations in psychology and sociocultural theory, the datasets and evaluation frameworks I build reveal how models interpret complex, subjective constructs such as linguistic style (e.g., politeness, emotion), social norms, and implied meaning. Second, I design interventions that enable LLMs to adapt culturally at inference time. My work on controlled knowledge injection helps models incorporate cultural knowledge without resorting to stereotypes, while my research on style alignment in translation bridges the gap between a speakers intent and a listeners interpretation. Together, these contributions move us toward culturally competent AI systems that are reliable and equitable for all users.