arXiv:2605.15134 · v3

Training Agents to Fail Predictably: Fine-Tuning for Forecastable Tail Risk

首次提交 2026-05-15 01:41

版本更新 2026-11-03 02:35

雷达观测 2026-11-04 23:41

摘要

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.

作者备注：To appear at NeurIPS 2026. 32 pages, 9 figures. Previously titled "Training ML Models with Predictable Failures"