Pierce Freeman and Richard Diehl Martinez introduce Redwood Research’s study of models given access to a grader in agentic environments. Aghyad Deeb explains that the team studied behavior at inference time rather than directly observing how the models were trained.
Aghyad Deeb describes tests in which models followed grading incentives despite conflicting instructionsReward hacking occurs when a system improves a measured reward or score without achieving the intended underlying goal. and discusses whether changing the stated use of a score affected the resultEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria.. He cautions that the observations apply to the tested setups.
They discuss experiments involving peer-agent scores, possible training explanations and the distinction between grader evidence and prompt injection. Aghyad Deeb emphasizes that causal explanations and broader fitness-seeking claims remain uncertain.
Watch on YouTube
