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.
Quantifying the risk of rare failures in language models, such as those triggered by adversarial distribution shifts or very large-scale deployments, requires estimating probabilities far too small for random sampling. While recent work has formalized Low Probability Estimation, existing pipelines remain fragile in the rarest regimes: estimators can suffer zero-estimate collapse or systematic bias, and standard evaluation losses can become unstable or poorly matched to asymmetric safety costs. In this work, we introduce Gradient Activation Adaptive Multi-Level Splitting (GA-AMLS), which adapts rare-event Monte Carlo methods to the continuous activation space of language models. Specifically, GA-AMLS uses a gradient-based MCMC kernel to navigate activation space, eliminating the zero-estimate collapse of input-space search and replacing the independence assumptions of prior activation-space estimators with conditional sampling under an explicit, heavier-tailed activation prior. We also propose the Shifted-Power Bregman (SPB) Loss, a proper scoring rule that remains finite for zero-estimates and offers tunable asymmetry between underestimation and overestimation penalties. Experiments on small transformer models reveal a bias-variance tradeoff: GA-AMLS achieves the lowest loss under symmetric evaluation, reducing average log-space squared error relative to the strongest baseline across model sizes, while methods with overestimation bias prevail under asymmetric penalties. Our findings highlight that estimator choice should be matched to deployment context. More broadly, our work establishes activation space as a tractable domain for rare-event estimation in language models, circumventing the brittleness of discrete input-space search.
Pre-deployment safety evaluations aim to inform the downstream risks of releasing a new AI model. Yet most evaluations provide limited evidence about how often undesired model behavior will occur in deployment: they generally have insufficient coverage, are unrepresentative, and are generally recognizable as tests. To address these concerns, we study a simple way to simulate a model deployment: starting from de-identified conversations from a previous model deployment, we hold fixed the initial conversation prefix and regenerate the next response using a candidate model. The resulting responses can then both be audited for novel misalignments and used to estimate the prevalence of model misbehavior before deployment. We evaluate deployment simulation across four GPT-5-series deployments, using registered, outcome-blinded predictions for GPT-5.4 and retrospective analyses of three earlier releases. We find that deployment simulation produces informative estimates of post-deployment misbehavior rates and outperforms baselines based on adversarially selected production data; its evaluation-awareness point estimates were also much closer to production traffic than those from traditional evaluations. We also identify the realism of tool resampling as a central challenge for further improving predictions and share results suggesting that this challenge is surmountable even in complex tool-use settings. Finally, we show that deployment simulation can be seeded from public chat datasets and remain informative about production misbehavior rates, suggesting a path for external researchers to run deployment-grounded evaluations without access to private production logs. Overall, deployment simulation helps evaluators forecast how language models will behave in the real world and supports more quantitative assessment of deployment risk.
Marcus Williams, Hannah Sheahan, Cameron Raymond +8
We document inverse scaling in LLMs on forecasting problems whose underlying time series exhibit superlinear growth and tail risk of regime change, a structure common in finance and epidemiology. On these tasks, more capable models produce worse distributional forecasts. The pattern appears on ForecastBench-Sim (FBSim), a contamination-free, simulated-world benchmark we release, in forecasting synthetic SIR epidemics with a matched linear control, and replicates in real-world datasets on COVID-19, measles, housing markets, and hyperinflation. A per-quantile decomposition shows the failure concentrates at the upper tail, which more capable models shift upward to track aggressive extrapolations of growth, while the lower tail stays put. A within-family study of Llama-3.1 shows that both model scale and post-training independently contribute to this effect. Domain knowledge does not reliably rescue calibration. This inverse scaling does not appear on single-threshold metrics common in LLM forecasting benchmarks, reversing the sign of the capability--accuracy relationship on identical outputs. Single-threshold scoring at conventional cutoffs misses the upper-tail cost; tail-inclusive scoring reverses the sign of the capability--accuracy relationship on the same outputs. We recommend that LLM forecasting evaluations use continuous (and unbounded) measures of accuracy alongside bounded binary threshold metrics.