An importance-sampling method that perturbs model weights estimates agent failure probabilities down to 1e-9.

The question safe deployment needs answered isn't whether a catastrophic trajectory can occur but how often. This method builds 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, it delivers over 800x compute-weighted efficiency against naive Monte Carlo below 1e-7. Implementation released. (arXiv 2609.24969)