Gradient Ascent Unlearning
An unlearning method that reduces the probability of a model producing specific undesirable behaviors by performing gradient ascent on the negated loss. It often faces a trade-off between incomplete unlearning and excessive unlearning.
Timeline (last 90 days)
The study compared standard methods, including supervised fine-tuning (SFT), GRPO, and gradient ascent unlearning.
Gradient ascent unlearning negates the loss corresponding to the target undesirable behavior and then performs gradient ascent, thereby reducing the probability of the model producing that behavior.
All Facts (2)
The study compared standard methods, including supervised fine-tuning (SFT), GRPO, and gradient ascent unlearning.
50%UnverifiedGradient ascent unlearning negates the loss corresponding to the target undesirable behavior and then performs gradient ascent, thereby reducing the probability of the model producing that behavior.
50%