When Equal Isn’t Fair: Mitigating Over-Normalization in Large Language Models (Student Abstract)

Bias in Large Language Models (LLMs) is increasingly addressed through fairness-oriented techniques. However, in some cases, these approaches may inadvertently remove genuine cultural differences between groups, leading to “over-normalization” or models losing important socio-cultural distinctions. In this work, we introduce OverNormEval, a benchmark designed to detect when an LLM exhibits such over-normalization. We further explore the use of Direct Preference Optimization (DPO) to mitigate over-normalization.

Authors

• Ravada Satyadev IIIT-Delhi
• Aditya Ganesh Kumar Singapore Institute of Technology
• Avinash Anand Indraprastha Institute of Information Technology, Delhi
• Rajiv Ratn Shah Indraprastha Institute of Information Technology, Delhi
• Zhengkui Wang Singapore Institute of Technology
• Mukesh Prasad University of Technology Sydney

Keywords

No keywords are indexed for this paper.