mechanistic study

mlp weight (1) neural network (1) large language model (1) adversarial input (1) mlp weight transformation (1) jailbreak defense (1) safety fine-tuning (1) mechanistic interpretability (1) preference optimization (1) preference alignment (1)