Efficient Inference for Noisy LLM-as-a-Judge Evaluation
This work addresses the issue of imperfect LLM judges for researchers and practitioners in AI evaluation, providing incremental improvements in statistical efficiency for bias correction methods.
The paper tackles the problem of noisy LLM-as-a-judge evaluations by systematically comparing two bias-correction approaches for estimating mean parameters, showing that prediction-powered inference can achieve strictly smaller asymptotic variance under certain conditions, as verified in simulations and real-data examples.
Large language models (LLMs) are increasingly used as automatic evaluators of generative AI outputs, a paradigm often referred to as "LLM-as-a-judge." In practice, LLM judges are imperfect predictions for the underlying truth and can exhibit systematic, non-random errors. Two main approaches have recently been proposed to address this issue: (i) direct measurementerror correction based on misclassification models such as Rogan-Gladen-style estimators, and (ii) surrogate-outcome approaches such as prediction-powered inference (PPI), which correct bias by calibrating prediction residuals on a small set of gold-standard human labels. In this paper, we systematically study the performance of these two approaches for estimating mean parameters (e.g., average benchmark scores or pairwise win rates). Leveraging tools from semiparametric efficiency theory, we unify the two classes of estimators by deriving explicit forms of efficient influence function (EIF)-based efficient estimators and characterize conditions under which PPI-style estimators attain strictly smaller asymptotic variance than measurement-error corrections. We verify our theoretical results in simulations and demonstrate the methods on real-data examples. We provide an implementation of the benchmarked methods and comparison utilities at https://github.com/yiqunchen/debias-llm-as-a-judge.