Reproduce “Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yisong Xiao, Aishan Liu, Siyuan Liang, Xianglong Liu, Dacheng Tao
Researchers from Beihang University, National University of Singapore, and Nanyang Technological University developed "Fairness Mediator" (FairMed), an inference-time framework designed to mitigate societal biases in Large Language Models by targeting stereotype associations within the MLP components. FairMed achieved average bias reductions of up to 84.42% and 80.36% on BBQ metrics while preserving the LLM's language understanding capabilities.
Key insights
- Stereotype associations are encoded as entity-information pairs within the Multi-Layer Perceptron (MLP) layers of transformer models.
- Different bias types are encoded distinctly across model layers, with some biases becoming more pronounced in deeper layers.
- Bias can be effectively neutralized by iteratively adjusting MLP activations using gradient guidance from a prober, without degrading overall model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.