A used as a judge should change its verdict only when the task outcome changes, not when the evidence is presented differently. But current models often fail this test, making them unreliable for real-world evaluation.
The researchers built GridWM-Judge, a based on simple grid-world simulations. It creates matched pairs of trajectories where either the presentation is varied or the outcome is changed (counterfactuals). This separates nuisance from causal changes.
They found that many models are sensitive to presentation changes and weak at detecting counterfactual differences. This matters because using such judges can lead to incorrect conclusions about robot success. The work advocates for evaluating judges on both stability and causal sensitivity.