E-valuator: Reliable Agent Verifiers with Sequential Hypothesis Testing
PDF
Link
Code
@article{sadhuka2025valuator,
title = {E-valuator: Reliable agent verifiers with sequential hypothesis testing},
author = {Sadhuka, Shuvom and Prinster, Drew and Fannjiang, Clara and Scalia, Gabriele and Berger, Bonnie and Regev, Aviv and Wang, Hanchen},
journal = {arXiv preprint arXiv:2512.03109},
year = {2025},
url = {https://arxiv.org/abs/2512.03109}
}
Citation
Sadhuka, S., Prinster, D., Fannjiang, C., Scalia, G., Berger, B., Regev, A., & Wang, H. (2025). E-valuator: Reliable Agent Verifiers with Sequential Hypothesis Testing. arXiv preprint arXiv:2512.03109.
Informal Summary
Converts verifier scores into decision rules with statistical guarantees by framing trajectory evaluation as a sequential hypothesis testing problem over e-processes, which enables continuous monitoring of agent actions. Improves false alarm control and statistical power across several datasets while allowing early termination of unsuccessful trajectories to conserve computation.