The field of Large Language Models (LLMs) is rapidly evolving, with a focus on improving evaluation methodologies and ensuring safety in real-world applications. Recent developments have centered around creating more robust and adaptive evaluation frameworks, such as those utilizing evolutionary or adversarial data augmentation. These approaches have shown promise in uncovering vulnerabilities and improving model generalization. Additionally, there is a growing emphasis on reality-oriented safety evaluations, which aim to assess LLMs in more realistic and dynamic scenarios. Researchers are also exploring innovative defense mechanisms, including adaptive reasoning and reinforcement learning-based methods, to enhance model robustness and safety. Noteworthy papers in this area include AutoEvoEval, which introduces a novel evolution-based evaluation framework for close-ended tasks, and ROSE, which proposes a reality-oriented safety evaluation framework using multi-objective reinforcement learning. OMS is also notable for its on-the-fly, multi-objective, self-reflective ad keyword generation via LLM agent.