Google DeepMind has recently unveiled a groundbreaking method for prompting AI responses named Optimization by PROmpting (OPRO). This method utilizes large language models (LLMs) as optimizers where the AI models work by attempting different prompts until they find one that comes closest to solving a particular task. This technique is described in a research paper and automates the trial and error process that a person would typically do by typing.
Unlike traditional methods that rely on formal mathematical definitions, OPRO defines optimization tasks in natural language. The researchers highlighted, “Instead of formally defining the optimization problem… we describe the optimization problem in natural language, then instruct the LLM to iteratively generate new solutions based on the problem description and the previously found solutions.” This adaptability allows the LLM to tackle a diverse range of problems by merely adjusting the problem description or adding specific instructions.
How OPRO Operates
The OPRO process commences with a “meta-prompt” input, which encompasses a natural language description of the task, examples of problems, placeholders for prompt instructions, and corresponding solutions. As the optimization unfolds, the LLM produces candidate solutions based on the problem description and prior solutions in the meta-prompt.
These candidates are then evaluated, assigned quality scores, and optimal solutions are integrated into the meta-prompt, enhancing the context for subsequent solution generation. This iterative mechanism persists until the model ceases to suggest superior solutions. The researchers emphasized the LLMs' capacity to comprehend natural language, enabling users to articulate their optimization tasks without formal specifications.
To ascertain OPRO's efficacy, the researchers tested it on renowned mathematical optimization challenges like linear regression and the traveling salesman problem. The results were promising, with LLMs accurately capturing optimization directions based on prior optimization trajectories provided in the meta-prompt.
Why AI Auto Prompting Is A Breakthrough
Furthermore, experiments revealed that prompt engineering can significantly influence a model's output. For instance, appending the phrase “let's think step by step” to a prompt can induce the model to delineate the necessary steps to address a problem, often leading to enhanced accuracy.
However, it's important to note that LLMs' responses are profoundly influenced by the prompt's format, and similar prompts can produce different outcomes.
AI that is able to auto-prompt brings benefits to the user while bringing the following benefits to LLMs:
- Improved efficiency: Auto prompting can help to improve the efficiency of AI systems by reducing the amount of manual input and intervention required. For example, an AI system that is able to automatically generate prompts for itself can reduce the need for human users to provide explicit instructions.
- Increased accuracy: AI auto prompting can also help to improve the accuracy of AI systems by providing them with more context and information. For example, an AI system that is able to generate prompts based on its own understanding of a task is less likely to make mistakes than an AI system that is only able to follow explicit instructions.
- Reduced bias: LLMs that can auto prompt could also help to reduce bias in AI systems by providing them with a more diverse range of inputs. For example, an AI system that is able to generate prompts based on its own understanding of a task is less likely to be biased towards certain types of data or examples.