DA-RAC: Distance-Aware Calibration of LLM Judges for Trustworthy AI Auditing

LLM judges can be unreliable evaluators: Current AI systems often use large language models to score whether other AI outputs are good or harmful, but these "judges" can be fooled by irrelevant examples, leading to false confidence in low-quality or potentially harmful outputs that shouldn't pass evaluation.

Context matters more than previously recognized: The paper demonstrates that which reference examples an LLM judge sees significantly affects its scoring—a problem called "context-induced miscalibration"—meaning the same AI output could be judged differently depending on what comparison examples are shown to the evaluator.

DA-RAC improves evaluation reliability: The proposed method reduces "false-pass risk" (harmful outputs incorrectly approved) by selecting reference examples that are similar to what's being evaluated, rather than using random or generic examples, making evaluations more trustworthy for high-stakes applications.

AI auditing requires transparency and contestability: The paper emphasizes that automated evaluation systems need to show their work—making the reference examples and decision-making process visible and open to challenge—rather than operating as black boxes, especially when evaluating AI systems that affect real people.

Better models alone aren't enough: Simply using more advanced language models won't solve evaluation problems; the paper argues that how we select and present reference examples to judges is equally critical for responsible AI deployment.