Reproduce “Mitigating Biases in Language Models via Bias Unlearning” 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
Dianqing Liu, Yi Liu, Guoqing Jin, Zhendong Mao
Researchers from the University of Science and Technology of China and People's Daily Online developed BiasUnlearn, a framework to mitigate social biases in large language models while preserving their core capabilities. The method reduces StereoSet scores by 11.91 to 15.98 points with minimal degradation to the Language Modeling Score and demonstrates that debiasing weights are transferable from base models to their fine-tuned variants.
Key insights
- Bias representations are deeply entrenched during pre-training and persist through fine-tuning phases, allowing debiasing weights learned on base models to be robustly transferred to their instruction-fine-tuned variants.
- A combination of stereotype forgetting via negative optimization and anti-stereotype retention with an adversarial component can effectively mitigate bias while minimally impacting a model's general language understanding and generation abilities.
- Preventing bias reversal is critical for robust debiasing, and mechanisms like adversarial forget sets and dynamic dataset swapping provide effective regularization against over-debiasing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.