Tool audits sensitivity of large language model evaluation judges
MAWILE: Multi-Axis Workbench for Inspecting LLM Evaluators
Artificial Intelligence
Summary
Evaluating AI language models often relies on other AI systems called judges, but these judges can be easily confused by small changes in responses or instructions. The authors created MAWILE, a tool that helps developers test how sensitive these judges are to different kinds of changes. It does this by making controlled tweaks and checking if the judge’s rating changes appropriately or stays consistent when it should. This helps improve the reliability of AI evaluation without needing expert-labeled answers.
What this means in practice
- •For machine learning engineers: Check and improve the reliability of AI model evaluation judges during development by systematically testing their sensitivity to input and instruction changes.
- •For quality assurance teams: Audit evaluation tools used for validating AI outputs to detect unintended fragilities without needing gold-standard labels.
Authors
Jackson Hassell, Farima Fatahi Bayat, Pouya Pezeshkpour, Estevam Hruschka
Abstract
Large language model (LLM) judges provide a flexible and scalable method for evaluating model and agent outputs, but their verdicts can be sensitive to incidental changes in the evaluated response, judge instructions, and scoring rubric. Existing systems examine important subsets of these failure modes, but auditing a configured judge requires testing both the judge instrument and the items it evaluates. We introduce MAWILE, a developer-facing workbench for auditing judge sensitivity across four surfaces: the judge prompt, judge rubric, target-system input, and target-system output. Given a user-supplied judge and representative evaluation items, MAWILE constructs and validates controlled perturbations, re-executes the judge, and localizes the resulting sensitivity. Each perturbation declares whether the verdict should remain invariant or change in a specified direction, allowing the same system to measure both robustness to irrelevant variations and sensitivity to meaningful changes. MAWILE audits binary, ordinal, and pairwise judges without requiring gold labels. The code for this tool is available at: github.com/megagonlabs/mawile-judge.