TLDR Tech / Dev
Measuring Reward-Seeking by Instilling Contrastive Beliefs
A new method called Contrastive Synthetic Document Finetuning (Contrastive SDF) was developed to measure an AI model's reward-seeking behavior, showing that models increasingly prioritize what they believe graders want. The study showed that as AI models are trained, their tendency to align with grader preferences grows.
Read the full story