More Gemini Features Coming to Google TV
Google TV integrates Gemini AI for image/video generation, enhances Google Photos, and adds a dedicated YouTube Shorts feed to the homepage.
Read moreWhen people develop new deep learning AI models, the majority of them rely on optimization algorithms or optimizers to ensure that the models have high accuracy. However, the most commonly used optimizer - derivative-based optimizer - encounters problems when dealing with real-world applications.
In a new paper, researchers from DeepMind propose a new approach: Optimization by Prompting (OPRO), which uses large language models (LLMs) as optimizers. The unique aspect of this approach is that it defines the optimization task in natural language instead of through formal mathematical definitions.
The researchers write, "Instead of formally defining the optimization problem and using programming solvers to derive update steps, we describe the optimization problem in natural language and instruct the LLM to iteratively generate new solutions based on the problem description and previously found solutions."
This technique is highly adaptable. By simply modifying the problem description or adding specific instructions, the LLM can be guided to solve a wide range of problems.
The researchers found that for small-scale optimization problems, the LLM can generate effective solutions simply by being prompted, sometimes even matching or surpassing the performance of expert-designed heuristic algorithms. However, the true potential of OPRO lies in its ability to achieve maximum accuracy from the model by optimizing the prompts given to the LLM.
How OPRO Works
The OPRO process begins with inputting a "meta-prompt." This meta-prompt includes a natural language description of a task, as well as some examples of problems, placeholder instructions, and corresponding solutions.
As the optimization process unfolds, the large language model (LLM) generates candidate solutions. These solutions are based on the problem description and the previous solutions contained in the meta-prompt.
Then, OPRO evaluates these candidate solutions and assigns a quality score to each solution. Excellent solutions and their scores are added to the meta-prompt, providing rich context for generating the next round of solutions. This iterative process continues until the model stops generating better solutions.
The researchers explain, "The main advantage of LLM in optimization is its ability to understand natural language, which allows people to describe their optimization tasks without the need for formal specifications."
This means that users can specify target metrics, such as "accuracy," while providing additional instructions. For example, they can request the model to generate solutions that are both concise and widely applicable.
OPRO also leverages the LLM's ability to detect contextual patterns. This enables the model to identify an optimization trajectory based on the examples contained in the meta-prompt. The researchers note, "Including optimization trajectories in the meta-prompt allows the LLM to recognize the similarity of high-scoring solutions, encouraging the LLM to build potentially better solutions based on existing good solutions without explicitly defining how to update the solutions."
To validate the effectiveness of OPRO, the researchers tested it on two well-known mathematical optimization problems: linear regression and the "traveling salesman problem." While OPRO may not be the optimal method for solving these problems, the results were encouraging.
The researchers report, "In both tasks, we observed that LLMs we evaluated were able to act as optimizers, consistently improving the performance of generated prompts throughout the iterative optimization process until convergence."
Optimizing LLM Prompts through OPRO
Experiments have shown that prompt engineering can greatly influence the output of the model. For example, appending the phrase "let's think step by step" to the prompt can guide the model to reason and outline the steps required to solve the problem, often resulting in more accurate results.
However, it is important to remember that this does not mean LLMs possess human-like reasoning abilities. Their responses are highly dependent on the format of the prompt, and semantically similar prompts can yield drastically different results. The DeepMind researchers write, "The optimal prompt format may be model-specific and task-specific."
The true potential of OPRO lies in its ability to optimize prompts for LLMs like OpenAI's ChatGPT and Google's PaLM. It can guide these models to find the best prompts to maximize task accuracy.
The researchers write, "OPRO enables LLMs to progressively generate new prompts, improving task accuracy through the optimization process, where the initial prompt has low task accuracy."
To illustrate this, consider the task of finding the best prompt to solve a word math problem. Given a meta-prompt containing instructions and examples with optimization prompts (e.g., "let's think step by step"), provide a meta-prompt for "optimizing LLM." The model generates a set of different optimization prompts and passes them to the "scoring LLM." The scoring LLM tests these prompts on the problem examples and evaluates the results. The best prompt and its score are added to the beginning of the meta-prompt, and this process is repeated.
The researchers evaluated this technique using several LLMs from the PaLM and GPT series. They found that "all the LLMs we evaluated were able to act as optimizers, consistently improving the performance of generated prompts throughout the iterative optimization process until convergence."
For example, when testing OPRO on PaLM-2 with GSM8K, a benchmark test based on school math word problems, the model produced interesting results. It started with the prompt "let's solve the problem" and generated other strings such as "let's think carefully about the problem and solve it together," "let's break down the problem," "let's calculate the solution," and finally "let's do the math calculations," providing the highest accuracy.
In another experiment, adding the string "take a deep breath and solve this problem step by step" resulted in the most accurate results.
These results are both fascinating and somewhat unsettling. For humans, all these instructions have the same meaning, but they trigger very different behaviors in LLMs. This reminds us not to attribute human-like characteristics to LLMs and highlights how much we still have to learn about their internal workings.
However, the advantages of OPRO are evident. It provides a systematic approach to explore the potential space of LLM prompts and find the best prompts for specific types of problems. Its performance in practical applications remains to be observed, but this research takes us one step closer to understanding how LLMs work.
Google TV integrates Gemini AI for image/video generation, enhances Google Photos, and adds a dedicated YouTube Shorts feed to the homepage.
Read moreAlphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Read moreAnthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Read moreApple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Read moreGoogle launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
Read moreYouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Read moreGoogle grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Read moreMistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.
Read moreCustom subliminal audio maker

AI Social Media Manager

AI Search Visibility

Workout and Calorie Tracker
Stock research platform

AI interview assistant

Scientific figure generator

AI courses with real code