Since ChatGPT dropped in the fall of 2022, prompt engineering has become a popular practice among users of large language models (LLMs) to enhance their results. Prompt engineering involves finding clever ways to phrase queries to LLMs to get the desired output. However, new research suggests that prompt engineering may not be as effective as previously thought. In fact, the study indicates that allowing the model to optimize its own prompts may yield better results.

Illustration: www.youtube.com

In a study conducted by researchers at VMware, it was found that prompt engineering techniques yielded inconsistent results. Different strategies, such as using chain-of-thought prompting or positive prompts, had varying effects on LLM performance. The researchers concluded that there is no one-size-fits-all approach to prompt engineering. What works for one model may not work for another, emphasizing the need for model-specific optimization.

The study also explored the possibility of having the model optimize its own prompts. New tools have been developed to automate this process, allowing the model to find the optimal phrase to feed into the LLM. The researchers found that in almost every case, the model-generated prompts outperformed the best human-engineered prompts. Additionally, the process of generating prompts algorithmically was much faster compared to manual optimization.

Illustration: www.frontiersin.org

Prompt engineering is not limited to language models but can also be applied to image-generation algorithms. A team at Intel Labs developed a tool called NeuroPrompts, which automatically enhances simple prompts to create visually stunning images. The tool utilizes reinforcement learning to optimize prompts based on aesthetic criteria. Once again, the automatically generated prompts surpassed the prompts created by human experts.