Automated Prompt Engineering

Automated Prompt Engineering – APE

The ascent of large language models (LLMs) marks a significant milestone in natural language processing (NLP) and related fields. Yet, unlocking their full potential hinges on effective prompt engineering—the art of crafting precise instructions that steer the model towards generating desired outputs. Automated Prompt Engineering (APE) emerges as a automated solution to the intricacies of prompt engineering, offering a systematic approach to generating prompts.

APE embodies a two-fold objective: enhancing the precision of outputs and curbing hallucinations, which are false or incorrect pieces of information generated by LLMs. Unlike the traditional manual approach, which requires an in-depth understanding of the domain and the model, APE automates prompt generation.

In this context, a distinct feature of APE is its treatment of instructions as "programs." This notion propels a search over a variety of instruction candidates, proposed by an LLM, to ascertain the most effective instruction that maximizes the accuracy of the generated outputs. By framing instructions as programs, APE introduces a computational rigor to prompt engineering, paving the way for more sophisticated interactions with LLMs.

The Mechanism Behind Automated Prompt Engineering

Delving into the mechanics of Automated Prompt Engineering unveils a fascinating orchestration between two large language models: a prompt generator and a content generator. The prompt generator is the maestro, crafting prompts based on the input it receives, while the content generator performs the symphony, producing outputs given the prompts.

The process commences with feeding a prompt alongside a small set of example input-output pairs to the prompt generator. A concrete example would be the prompt: “I gave a friend an instruction and five inputs. The friend read the instruction and wrote an output for every one of the inputs. Here are the input-output pairs:” followed by a set of example inputs and outputs. Post this, the prompt generator, leveraging its trained prowess, constructs a new prompt, such as “Choose the animal that is bigger” based on the provided context. This newly minted prompt is then fed along with example inputs from a dataset to the content generator, which, in turn, generates outputs.

The generated outputs are then scrutinized for quality based on their alignment with the expected outputs. The prompt’s effectiveness is gauged by how often the content generator produces outputs that exactly match the expected outputs. To further sharpen the prompt, the prompt generator is asked to produce a variation of the highest-scoring prompt, and the process is iterated to refine the prompt to its optimum form.

Optimizing prompt generation

Automated Prompt Engineering operates through a nuanced mechanism that seamlessly intertwines different modes of generation and customization to optimize prompt generation. Here’s a simplified elucidation of these critical components:

Forward Mode Generation: In this mode, APE aims to generate high-quality instruction candidates by translating a specific distribution into words. Essentially, it follows a left-to-right text generation approach, akin to reading a book from the start to the end. This mode is particularly effective when the instruction is positioned at the end of the prompt, following the natural flow of text generation.

Reverse Mode Generation: Contrarily, the Reverse Mode embarks on a more flexible path. It employs advanced LLMs capable of infilling, meaning they can fill in the missing instructions within a text, irrespective of the position. This mode is instrumental when the instruction needs to be placed anywhere other than the end, providing a more versatile approach to instruction generation.

Customized Prompts: APE also accommodates the customization of prompts based on the specific score function being utilized. This feature is particularly beneficial in experiments where human-designed instructions serve as the starting point, and APE’s Reverse Model is employed to propose initial instruction samples, thereby fitting the missing context aptly.


The emergence of prompt engineering is a relatively recent phenomenon, and yet, the introduction of Automated Prompt Engineering illustrates a swift progression towards automation in this domain. This transition reflects the broader trend within the AI landscape, where manual processes are continually being reviewed for automation to enhance efficiency and scalability.

Unlock the Future of Business with AI

Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.

Scroll to top