Aligning large language models (LLMs) with human values and expectations has traditionally been a costly and labor-intensive process, involving techniques like supervised fine-tuning and reinforcement learning from human feedback (RLHF). But what if AI could learn to align itself? New research introduces Dynamic Rewarding with Prompt Optimization (DRPO), a groundbreaking method that allows LLMs to self-improve without any additional training or human intervention. This tuning-free approach leverages a search-based optimization framework where the LLM iteratively refines its own alignment instructions. At the heart of DRPO lies a dynamic rewarding mechanism. This mechanism identifies and corrects the model's unique alignment weaknesses, enabling it to adapt to various alignment challenges. Imagine an LLM with outdated information attempting to answer a question requiring current events knowledge. DRPO would recognize this deficiency and adjust the alignment prompt to address it. This approach also optimizes in-context learning examples, which have been shown to be effective in guiding LLM behavior. Experiments across several LLMs, including Mistral, Llama 2, GPT-3.5 Turbo, and GPT-4, have revealed that DRPO substantially enhances alignment, even allowing base models to surpass their fine-tuned counterparts. Interestingly, prompts automatically optimized by DRPO outperformed those manually crafted by humans. This discovery indicates that current LLMs possess remarkable potential for self-alignment, paving the way for more adaptable and personalized AI systems. While promising, DRPO faces challenges, including the computational cost of optimization, the risk of imprecise reward assignments, and the limitations of LLMs' self-correction abilities. Despite these challenges, DRPO represents a significant step towards more autonomous and efficient LLM alignment, opening exciting avenues for future research.