Reproduce “Mitigating Social Bias in Large Language Models: A Multi-Objective Approach within a Multi-Agent Framework” 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
Zhenjie Xu, Wenqing Chen, Yi Tang, Xuanying Li, Cheng Hu, Zhixuan Chu, Kui Ren, Zibin Zheng and 1 more
MOMA treats social-bias mitigation as a joint fairness–accuracy optimization problem for black-box language models. It uses separate assistant agents to mask or rebalance identity-related cues before a task agent answers, reducing BBQ bias scores by up to 87.7% and improving StereoSet intrasentence ICAT from 66.83 to 86.99 for Llama-3, with task-performance losses generally below 6.8%.
Key insights
- Fairness prompting is more effective when applied as a targeted intervention on social-group representations rather than as a broad instruction to answer ethically.
- Balancing must be relational: adding generally positive adjectives can worsen disparities unless attributes are distributed to counter stereotypical associations between groups.
- Multi-agent orchestration alone does not ensure debiasing; the benefit comes from assigning agents distinct roles for input transformation and answer generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.