Training Deliberative Monitors for Black-Box Scheming Detection

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Sinha, Aditya, Naik, Akshat, Gillioz, Victor, Storf, Simon, Merkelbach, Kilian, Barton-Cooper, Rich, Højmark, Axel, Hobbhahn, Marius
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914612956364800
author Sinha, Aditya
Naik, Akshat
Gillioz, Victor
Storf, Simon
Merkelbach, Kilian
Barton-Cooper, Rich
Højmark, Axel
Hobbhahn, Marius
author_facet Sinha, Aditya
Naik, Akshat
Gillioz, Victor
Storf, Simon
Merkelbach, Kilian
Barton-Cooper, Rich
Højmark, Axel
Hobbhahn, Marius
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_29601
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Training Deliberative Monitors for Black-Box Scheming Detection
Sinha, Aditya
Naik, Akshat
Gillioz, Victor
Storf, Simon
Merkelbach, Kilian
Barton-Cooper, Rich
Højmark, Axel
Hobbhahn, Marius
Computation and Language
Artificial Intelligence
Machine Learning
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.
title Training Deliberative Monitors for Black-Box Scheming Detection
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2605.29601