UDA: Unsupervised Debiasing Alignment for Pair-wise LLM-as-a-Judge