An Importance-Sampling Method That Perturbs Model Weights Estimates Agent Failure Probabilities Down to 1e-9

arXiv·medium signalXBlueskyLinkedInCopy link

"Rare Event Estimation via Iterative Unalignment" (arXiv 2609.24969, 21 Sep 2026) targets the question safe deployment actually needs answered: not whether a catastrophic agent trajectory can occur but how often. The method constructs the importance-sampling proposal by perturbing the original model's weights, making the proposal itself a differentiably parameterized language model so the search runs by gradient descent over weight space, with an adaptive regularizer trading event amplification against estimator stability. On ~120M and ~2.6B models across three event families and 300+ rare events down to 1e-9, the estimator delivers over 800x compute-weighted efficiency versus naive Monte Carlo below 1e-7, with implementation released.