How Google DeepMind’s OPRO Transforms LLMs into Problem-Solving Tools

Why Trust Techopedia
Key Takeaways

In recent years, Large Language Models (LLMs) have evolved with enhanced text generation abilities, including in-context learning. Google DeepMind's OPRO empowers LLMs as versatile optimization tools, resembling emergent behavior in various fields. Effective prompts, like Chain-of-Thought, lead to new ways of problem solving via AI.

In recent years, there has been a concerted effort to scale up language models into what we now call Large Language Models (LLMs), which involves training larger models on more extensive datasets with increased computational power — resulting in consistent and expected improvements in their text generation abilities.

As LLMs continue to grow, they reach a point at which they unlock new capabilities, a phenomenon known as in-context learning or prompt-based learning.

These newfound skills develop naturally without specific training, enabling LLMs to perform tasks such as arithmetic, answering questions, and summarizing text, all acquired through exposure to natural language.

This excitement has recently taken on a new dimension as researchers from Google DeepMind have transformed LLMs into powerful optimization tools using their prompting technique, known as Optimization by PROmpting (OPRO).

In-context or Prompt-based Learning: An Emergent Behavior of LLMs

Emergent behavior describes how a system can drastically change its behavior when minor adjustments are made within it, especially as it reaches a specific threshold.

A prime example of emergent behavior can be seen in water. As the temperature decreases, the behavior of water gradually changes, but there’s a critical point where something remarkable happens. At this specific temperature, water undergoes a rapid and significant transformation, transitioning from a liquid state to ice, much like flipping a switch.


Emergent behavior is not limited to specific fields but across various domains such as physics, biology, economics, and systems. In the context of LLMs, however, this means that after a particular stage in their training, LLMs appear to transition into a new mode where they can effectively tackle complex problems without explicit training.

This remarkable behavior is usually initiated and guided using prompts, which are natural language instructions provided to the LLMs. Because the quality of LLM responses is closely tied to the quality of the prompt, crafting effective prompts has evolved into a pivotal element of LLM utilization.

For example, Chain-of-Thought is a prompting technique developed to enable the model to break down complex problems into sub-problems and chaining them together to solve problems in a way we solve mathematical and reasoning problems. This behavior is achieved by providing both the intermediate reasoning steps and the final solution as a prompt to guide LLMs to accomplish these tasks.

For example, to enable the LLM to solve common sense reasoning tasks like “I’m going on a hike and need to pack water. How many 16-ounce water bottles should I bring for a 10-mile hike?”, we can prompt the model like “A general guideline is to drink about 0.5-1 liter (17-34 ounces) of water per hour of hiking. For a 10-mile hike, you’d want at least 1-2 bottles, so two bottles of 16 ounces each should be sufficient.”

Evolution of LLMs into Powerful Optimizers

Contemporary AI research is witnessing a burgeoning interest in developing innovative techniques for effectively prompting LLMs, leveraging their emergent capabilities to tackle problem-solving tasks.

In this context, researchers at Google DeepMind have recently achieved a significant breakthrough with a new prompting technique known as “Optimization by PROmpting” (OPRO), which can prompt LLMs to solve optimization problems. This emergent optimization ability adds a new layer of utility to these LLMs, making them valuable problem-solving tools in various domains.

Consider the possibilities. You can present a complex engineering problem in plain English rather than formally defining the problem and deriving the update step with a programmed solver. The language model can grasp the intricacies and propose optimized solutions. Similarly, financial analysis can assist in portfolio optimization or risk management. The applications span a broad spectrum, from supply chain management and logistics to scientific research and creative fields like art and design.

How Does OPRO Work?

In a nutshell, OPRO uses the power of language models to solve problems by generating and evaluating solutions, all while understanding regular language and learning from what it’s done before. It’s like having a clever assistant that keeps getting better at finding solutions as it goes along. An essential component of this process is meta-prompt, which has two key parts:

• First, it explains the problem in words, including what we’re trying to achieve and any rules we must follow. For example, if we’re trying to improve the accuracy of a task, the instructions might say “come up with a new way to make the task more accurate.”

• Second, it includes a list of solutions the LLM had tried before and how good they were. This list helps the LLM recognize patterns in the answers and build on the ones that seem promising.

During each step of the optimization process, the LLM comes up with potential solutions for the optimization task. It does this by considering both the problem description and the solutions it has seen and evaluated before, which are stored in the meta-prompt.

Once it generates these new solutions, they are carefully examined to see how good they are at solving the problem. They are added to the meta-prompt if they outperform the previously known solutions. This becomes a cycle where the LLM keeps improving its solutions based on its learning.

To understand the idea, consider the task of optimizing a financial portfolio. An “optimizer LLM” is provided with a meta-prompt containing investment parameters and examples with placeholders for optimization prompts. It generates diverse portfolio allocations. These portfolios are evaluated by a “performance analyzer LLM” based on returns, risk, and other financial metrics. The prompts for the highest-performing portfolios and their performance metrics are integrated into the original meta-prompt. This refined meta-prompt is then used to improve the initial portfolio, and the cycle repeats to optimize investment results.

The Bottom Line

Advancements like OPRO are a paradox—captivating in their boundless potential to expand our horizons and disconcerting as they usher in an era where AI can autonomously craft intricate processes, including optimization, blurring the lines of human control and creation.

Nevertheless, the ability to transform Large Language Models (LLMs) into powerful optimizers establishes OPRO as a robust and versatile approach to problem-solving. OPRO’s potential spans engineering, finance, supply chain management, and more, offering efficient, innovative solutions. It marks a significant step in AI’s evolution, empowering LLMs to continuously learn and improve and opening new possibilities for problem-solving.


Related Reading

Related Terms

Dr. Tehseen Zia
Tenured Associate Professor
Dr. Tehseen Zia
Tenured Associate Professor

Dr. Tehseen Zia has Doctorate and more than 10 years of post-Doctorate research experience in Artificial Intelligence (AI). He is Tenured Associate Professor and leads AI research at Comsats University Islamabad, and co-principle investigator in National Center of Artificial Intelligence Pakistan. In the past, he has worked as research consultant on European Union funded AI project Dream4cars.