Ghosts in the Model
Machine unlearning and interpretability study showing that forgotten behavior can be restored through activation patching.
Ghosts in the Model investigates machine unlearning: whether unlearning methods genuinely erase knowledge from models or suppress behavior while leaving recoverable internal representations behind.
- Evaluated Gradient Ascent and NegLoRA on GPT-2 over 59 landmark-to-city forget prompts.
- Reduced forget probability to 7.6% and 0.8% across the two methods.
- Designed Lazarus activation patching to restore 89.7-100% of forgotten behavior.
- Used probing and causal analysis to show that knowledge remained decodable even after behavioral suppression.