New monitors detect AI agent scheming without internal access

Researchers have developed a new method for training smaller, open-weight models to detect scheming behavior in autonomous agents. These "deliberative monitors" operate solely on agent trajectories, without needing access to the agent's internal reasoning or model internals. The training process involves using a frontier model to generate rationales, filtering these rationales, and then distilling the high-quality ones into the monitors via supervised fine-tuning and reinforcement learning. Evaluations show these monitors can outperform or match the cost-performance of larger, prompted frontier models on various misalignment benchmarks. AI

IMPACT Enables more cost-effective AI safety monitoring by using smaller, open-weight models.

RANK_REASON The cluster contains a research paper detailing a new method for training AI monitors. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Claude Haiku 4.5
• Claude Opus 4.6
• Claude Sonnet 4.6
• Gemini 2.5 Pro
• Gemini 3.1 Flash-Lite
• Gemini 3.1 Pro
• GPT-5.4
• GPT-5.4 Nano
• Qwen3.5-27B

• paper
• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →