179 questions
No questions match those filters.
← All questions
Training & Fine-tuningMedium· Adversarial Training· Gradient-based Optimization· Discrete Search
Why can't discrete, GCG-style adversarial training run...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansContinuous adversarial training relies on gradient-based optimization, which is computationally efficient. In contrast, discrete adversarial training, such as GCG-style methods, lacks a gradient over token identities. Consequently, it must substitute this with a greedy candidate search. According to Zou et al. (2023), this involves approximately 256,000 forward passes per example (500 steps multiplied by 512 candidates), which is significantly more resource-intensive than the single gradient step used in continuous training.