Reproduce “UDA: Unsupervised Debiasing Alignment for Pair-wise LLM-as-a-Judge” on OpenResearch

About this paper

Yang Zhang, Cunxiang Wang, Lindong Wu, Wenbo Yu, Yidong Wang, Guangsheng Bao, Jie Tang

A framework named UDA was developed to mitigate preference biases in LLM-as-a-judge evaluations, dynamically adjusting Elo ratings using a neural network trained on unsupervised consensus signals and judge self-awareness features. This approach reduced inter-judge disagreement by over 60% and increased alignment with human preferences by 24.7% in zero-shot evaluations.

Key insights

• LLM judges exhibit heterogeneous preference biases, including self-preference, resulting in significant inconsistencies in evaluation scores across different models.
• The collective consensus ranking derived from a diverse pool of LLM judges can serve as an effective, unsupervised proxy target to reduce aggregate system variance and improve evaluation stability.
• Integrating 'self-awareness' features, which quantify a judge's own stylistic preferences, is critical for meaningful debiasing, preventing convergence to a 'false consensus' that deviates from human judgment.

Summary by alphaXiv.