Training ML Models with Predictable Failures
This work addresses the critical problem of pre-deployment safety assessment for ML models by providing a practical method to estimate rare failures, though the experiments are limited to small-scale proof-of-concept settings.
Jones et al. (2025) propose a method to predict deployment-scale failure rates of ML models by extrapolating from the largest k failure scores in an evaluation set, and introduce a forecastability loss fine-tuning objective that reduces held-out forecast error by 40-60% in two proof-of-concept experiments while preserving primary-task performance.
Estimating how often an ML model will fail at deployment scale is central to pre-deployment safety assessment, but a feasible evaluation set is rarely large enough to observe the failures that matter. Jones et al. (2025) address this by extrapolating from the largest k failure scores in an evaluation set to predict deployment-scale failure rates. We give a finite-k decomposition of this estimator's forecast error and show that it has a built-in bias toward over-prediction in the typical case, which is the safety-favorable direction. This bias is offset when the evaluation set misses a rare high-failure mode that the deployment set contains, leaving the forecast to under-predict at deployment scale. We propose a fine-tuning objective, the forecastability loss, that addresses this failure mode. In two proof-of-concept experiments, a language-model password game and an RL gridworld, fine-tuning substantially reduces held-out forecast error while preserving primary-task capability and achieving safety similar to that of supervised baselines.