Baker, 2025 — Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
Finds that a weaker model reading a reasoning model's chain of thought catches reward hacking better than action monitoring, and that strong optimization pressure against flagged reasoning teaches obfuscated reward hacking.
Publication links
Finds that a weaker model reading a reasoning model’s chain of thought catches reward hacking better than action monitoring, and that strong optimization pressure against flagged reasoning teaches obfuscated reward hacking.