arXiv:2603.00829v2 Announce Type: replace-cross 
Abstract: Safe deployment of Large Language Model (LLM) agents in autonomous settings requires reliable oversight mechanisms. A central challenge is detecting scheming, where agents covertly pursue misaligned goals. One approach to mitigating such risks is LLM-based monitoring: using language models to examine agent behaviors for suspicious actions. We study constitutional black-box monitors: prompted classifiers that detect scheming using only externally observable inputs and outputs, optimized on synthetic data generated from natural-language behavior specifications. We introduce two pipelines for generating synthetic agent trajectories, STRIDE (iterative refinement) and Gloom (agent-environment simulation), from which we generate 1,000 samples each. We optimize frontier LLM monitors on these datasets via prompt sweeps, human refinement, and automated prompt optimization, and evaluate performance on 7,500 held-out trajectories from ControlArena, a suite of grounded environments where agents operate in more realistic contexts. Our results demonstrate that monitors selected purely on synthetic data can generalize to more realistic environments, capturing a meaningful scheming signal. However, we find that performance saturates quickly in our setting, with simple prompt sweeps matching the results of more extensive optimization. Pushing beyond this limit yields no further improvements and instead leads to overfitting.

قدمت الأبحاث الأخيرة نهجًا لمراقبة الصندوق الأسود الدستوري لوكلاء نماذج اللغة الكبيرة (LLM)، مع التركيز على اكتشاف السلوكيات الماكرة حيث يسعى الوكلاء لتحقيق أهداف غير متوافقة. تستخدم هذه الطريقة المراقبة المعتمدة على LLM لتحليل تصرفات الوكلاء من خلال المدخلات والمخرجات القابلة للملاحظة خارجيًا، مع تحسين الأداء على بيانات اصطناعية تم إنشاؤها من مواصفات سلوك اللغة الطبيعية.

Investigaciones recientes han introducido un enfoque de monitoreo constitucional en caja negra para agentes de Modelos de Lenguaje Grande (LLM), centrado en detectar comportamientos engañosos donde los agentes persiguen objetivos desalineados. Este método utiliza el monitoreo basado en LLM para analizar las acciones de los agentes a través de entradas y salidas observables externamente, optimizando el rendimiento en datos sintéticos generados a partir de especificaciones de comportamiento en lenguaje natural.

Des recherches récentes ont introduit une approche de surveillance constitutionnelle en boîte noire pour les agents de modèles de langage de grande taille (LLM), axée sur la détection des comportements sournois où les agents poursuivent des objectifs mal alignés. Cette méthode utilise la surveillance basée sur LLM pour analyser les actions des agents à travers des entrées et sorties observables de l'extérieur, optimisant les performances sur des données synthétiques générées à partir de spécifications de comportement en langage naturel.

Recent research has introduced a constitutional black-box monitoring approach for Large Language Model (LLM) agents, focusing on detecting scheming behavior where agents pursue misaligned goals. This method utilizes LLM-based monitoring to analyze agent actions through externally observable inputs and outputs, optimizing performance on synthetic data generated from natural-language behavior specifications.