Newest papers

Simulated Control Evals

Jeffrey Olmo, David Zhang, Tyler Tracy · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.31593 · cited by 0 works (OpenAlex) · open-access PDF

French

Les évaluations de contrôle de la sécurité baschespécifique de base de Linux mesurent si les protections peuvent empêcher un modèle d'IA de causer des résultats indésirables même lorsqu'il essaie intentionnellement de le faire. Ces évaluations utilisent généralement des paramètres de contrôle de tâche conçus à des fins et extrapolent leurs résultats vers le déploiement d'un modèle. Cependant, l'efficacité des protections dans le déploiement peut dépendre des caractéristiques des environnements et des tâches que les paramètres de contrôle ne représentent pas. Pour élargir l'endroit où les évaluations de contrôle peuvent être exécutées, nous introduisons des évaluations de contrôle de simulation (SCEs), qui utilisent un modèle de langue pour simuler les répons

Machine translation (Meta M2M100); the English original follows.

Abstract

Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. These evaluations typically use purpose-built control settings and extrapolate their results to a model's deployment. However, the effectiveness of safeguards in deployment can depend on features of environments and tasks that control settings do not represent. To broaden where control evaluations can be run, we introduce simulated control evaluations (SCEs), which use a language model to simulate tool responses in place of an environment. We evaluate SCEs on LinuxArena, where execution provides ground-truth outcomes. We compare SCE prediction errors with those from a separate control setting, BashArena, and with variation in control metrics within LinuxArena. For independently sampled evaluation and deployment task sets, SCEs predict safety with a root mean squared error of 3.6 percentage points across five monitors at a 5% audit budget, compared with 10.1 for BashArena. SCEs closely reproduce distributions of attack strategies, attacker reasoning, and monitor reasoning, with over tenfold lower KL divergence from execution than BashArena. SCEs explain 78-85% of the between-environment variance in attack success, monitor scores, and safety. This accuracy is sample efficient: SCEs predict safety with less error than BashArena using fewer than 5% as many task-pair draws as the deployment task set. Strategies discovered through simulation also improve monitor performance on held-out LinuxArena attacks, outperforming BashArena and approximately matching improvement from executed strategies. These results suggest that deployment-specific SCEs can be more predictive than control evaluations executed in separate control settings.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Software deployment, Computer science, Task (project management), Control (management), Variance (accounting), Matching (statistics)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange