Predicting an agent's final outcome from its early trajectory promises to reduce evaluation costs by terminating runs once the result is sufficiently certain. This approach hinges on the predictor's confidence being well-calibrated—that is, a stated 80% confidence should match an 80% actual success rate. The new arXiv preprint (2609.25647) conducts a testing-driven reliability audit of such predictors.
The audit finds that calibration is target-specific: a predictor calibrated on one task or target does not retain its calibration when applied to another target, even within the same benchmark. This means that confidence scores can mislead when the predictor is used across different targets, potentially causing premature termination with incorrect outcomes.
The authors emphasize the need for rigorous reliability testing before relying on early outcome prediction in agent evaluation. Without such audits, cost savings from early stopping could come at the expense of accuracy. The study underscores that calibration transfer cannot be assumed, even in controlled single-benchmark settings.