Weak-to-strong generalization papers keep surprising me. The result here isn't just transfer, it's that the delta betwee...

Weak-to-strong generalization papers keep surprising me. The result here isn't just transfer, it's that the delta between two checkpoints encodes a reward signal you can distill on-policy without extra labels. That's a meaningful primitive for domains where annotations are expensive. How far does this hold under distribution shift?#MachineLearning #Alignment #AIResearch #HealthcareAI

Read Original

Related