Search papers, labs, and topics across Lattice.
This paper introduces a method for training action-only deliberative monitors to detect scheming behavior in black-box autonomous agents, using a scheming specification to elicit rationales from a frontier teacher, filtering them with a judge, and distilling them into open-weight monitors via supervised fine-tuning and RL. The approach enables effective scheming detection without relying on internal agent information or expensive frontier models. Experiments across five datasets and six misalignment benchmarks demonstrate that monitors trained with this method outperform low-cost prompted frontier models and even match the performance of more powerful models at a fraction of the inference cost.
Open-weight monitors, trained via rationale distillation, can detect AI scheming as well as or better than expensive, closed-source frontier models, offering a practical and transparent alternative for AI safety.
As autonomous agents become more capable of performing real-world tasks, distinguishing scheming behavior from benign task pursuit may become a central AI control problem. Existing monitors often rely on chain-of-thought access or internal activations, or use prompted frontier models, all of which can be unavailable, unreliable or expensive in deployment. In this work, we study action-only deliberative monitors: smaller open-weight models trained to detect scheming and sabotage from agentic trajectories without accessing the monitored agent's reasoning or model internals. Our method, inspired by deliberative alignment, uses a scheming specification to elicit structured rationales from a frontier teacher, filters them with a separate judge, and distills the highest-quality rationales into open-weight monitors with supervised fine-tuning and reinforcement learning. We train on five datasets, and evaluate across six out-of-distribution agentic misalignment benchmarks. We show that applying our method to Qwen3.5-27B yields higher performance than all low-cost frontier models as prompted monitors (Gemini 3.1 Flash-Lite, GPT-5.4 Nano, and Claude Haiku 4.5) and than Gemini 2.5 Pro, while also achieving lower marginal inference cost (token-metered USD per 1,000 evaluations). Stronger prompted frontier monitors (Gemini 3.1 Pro, GPT-5.4, Claude Sonnet 4.6, and Claude Opus 4.6) achieve higher performance but at roughly $16$--$34\times$ higher marginal inference cost. Several of our trained monitors are positioned on the empirical cost--performance Pareto frontier among the monitors we evaluate, providing practical low-cost, low-FPR alternatives to prompted frontier models.
