As language models are becoming widely adopted, it is important to ensure that they do not perpetuate harmful stereotypes. This project investigates model editing—a technique for modifying a model’s knowledge of a specific fact without affecting its broader knowledge—as a potential debiasing tool. We examine how debiasing model edits alter the underlying features of a target group. Specifically, we decompose the features that distinguish a target group (e.g. Asian women) from its corresponding broader concept (e.g. people) in the model’s embedding space and track how these features shift pre- and post-edit. Our findings shed light on the robustness of model editing as a cost-effective debiasing strategy and provide insight into how edits transform internal representations.

By participating in SuperUROP, I hope to gain more hands-on experience completing an end-to-end research project. I’m particularly excited to apply my prior research experience in mechanistic interpretability from UROPs and summer programs to debiasing research. I look forward to learning more about interpretability and bias in AI while making a meaningful contribution in this space.