This is the pre-proceedings for the RLC 2026. You may expect minor changes.
Reinforcement Learning Journal, vol. 7, 2026, pp. TBD.
Will be presented at the Reinforcement Learning Conference (RLC), Montréal, Quebec, Canada, August 15–17, 2026.
Keywords: Conformal Prediction, Sequential Decision-Making, AI Safety.
Pretrained policies are rarely perfect, and their failures can be costly. We present a conformal wrapper for pretrained policies that preemptively issues warnings before failures with a user-specified lead-time, and we prove that unpreempted failure rate stays below any user-specified threshold. We also present a training objective for the conformal wrapper that achieves near-optimal false alarm rates in small tabular Markov processes (where computing the optimal rate is tractable). This training objective is empirically validated in more complex simulation environments, including humanoid locomotion and quadcopter obstacle avoidance. Similar to standard conformal prediction, the guarantees are policy-agnostic, and hold as long as calibration and testing episodes are exchangeable. We also experimentally probe performance when exchangeability is violated, and observe relatively graceful degradation of unpreempted failure rate.
Garrett Ethan Katz, Adebayo Braimah, Qinru Qiu, and Simon Khan. "Conformal Preemption of Failures in Sequential Decision-Making Agents." Reinforcement Learning Journal, vol. 7, 2026, pp. TBD.
BibTeX:@article{katz2026conformal,
title={Conformal Preemption of Failures in Sequential Decision-Making Agents},
author={Garrett Ethan Katz and Adebayo Braimah and Qinru Qiu and Simon Khan},
journal={Reinforcement Learning Journal},
volume={7},
pages={},
year={2026}
}