Measuring Reward-Seeking via Contrastive Belief Updates
Measures reward-seeking in OpenAI o3 RL checkpoints using synthetic document finetuning to set conflicting grader beliefs; finds reward-seeking rises through production RL training.
Research / Sources
Apollo Research
Understanding and evaluating deceptive AI. apolloresearch.ai
Focus All
Topic Alignment
Date Any time
Stress Testing Deliberative Alignment for Anti-Scheming Training
Apollo and OpenAI stress-test frontier models for covert scheming and find deliberative alignment anti-scheming training reduces but does not eliminate it, with models increasingly recognising evaluations.