Alignment Training Dynamics
Investigates how fine-tuning on misaligned data or from partially misaligned states affects model behavior and long-term alignment.
3 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.