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
Apollo Research
Understanding and evaluating deceptive AI. apolloresearch.ai
Focus All
Topic Evals + Alignment
Source Apollo Research · 1
Date Past 6 months
1 item