LLM-as-Judge on a Budget

LLM evaluation is computationally expensive and unreliable: Because LLM judges produce inconsistent results when scoring the same response multiple times, practitioners must query them repeatedly to get accurate scores—a costly process that limits how many model outputs can be fairly evaluated.

Smart resource allocation beats uniform approaches: The paper demonstrates that instead of asking an LLM judge the same number of questions about each response, allocating more queries to responses with higher uncertainty produces significantly better evaluation accuracy with the same total budget—a finding with direct implications for scaling AI safety evaluations.

Theoretical guarantees enable principled evaluation: The authors provide mathematical proof that their variance-adaptive method achieves near-optimal error rates of $\tilde{O}\left(\sqrt{\frac{\sum_{i=1}^K σ_i^2}{B}}\right)$, offering practitioners a principled framework rather than ad-hoc approaches for deciding how many times to query their evaluation systems.

Practical improvements demonstrated on real benchmarks: Testing on established datasets (Summarize-From-Feedback and HelpSteer2) shows the method substantially reduces estimation errors compared to standard uniform allocation, suggesting immediate applicability to current AI alignment and safety assessment workflows.

Budget constraints are a real bottleneck for responsible AI: This work highlights that computational costs of rigorous LLM evaluation can limit how thoroughly AI systems are assessed before deployment, making efficient evaluation methods essential for scaling responsible AI practices.