Improved Generation of Adversarial Examples Against Safety-aligned LLMs