DeepMind’s new timekeeping system helps improve performance in LLMs

Increasing the inference time is one of them the main topics of artificial intelligence in 2025and AI labs are attacking it from different angles. In its latest paper, Google DeepMind introduced the concept of “Mind Evolution,” a method that prepares large-scale solutions (LLMs) for planning and project planning.

Time-enhanced methods attempt to improve the performance of LLMs by allowing them to “think” more when presenting their answers. Basically, this means that instead of generating its solution all at once, the model is allowed to generate multiple solutions, iterate and improve its solutions, and explore different solutions to the problem.

LLM solutions are changing

Mind Evolution relies on two key components: research and genetic algorithms. Search algorithms are a a common component in many chronology methods. They allow LLMs to find the best way to find a solution. Genetic algorithms are powered by natural selection. They create and modify the responses of many people to achieve a goal, which is often referred to as “fitness work.”

Mind Evolution begins by creating a number of responses that are expressed in natural language. Solutions are developed by the LLM assigned to the problem description along with relevant information and instructions. Then the LLM institute evaluates each candidate and corrects them if they do not meet the criteria for solving the problem.

The algorithm then selects the parents of the next generation of answers by sampling from the existing population, with the top answers having a chance of being selected. It then generates new solutions through crossover (selecting parent pairs and combining their elements to create a new solution) and mutation (making random changes to the newly generated solutions). It also uses an evaluation method to evaluate new solutions.

The cycle of evaluation, selection and recombination continues until the algorithm reaches the correct path or completes more iterations.

One of the most important aspects of Mind Evolution is the work of analysis. Analysis methods testers often want the problem to be formulated from natural language into a stable, symbolic representation that can be processed by a solution program. Solving a problem can require significant expertise and a deep understanding of the problem to identify all the elements that need to be symbolized and how they relate to each other, which limits its usability.

In Mind Evolution, exercise is designed for natural language processing tasks where the answers are expressed in natural language. This helps prevent the system from setting up problems, as long as a solution tester is available. They also provide textual comments in addition to numerical data, which allows the LLM to understand specific issues and make necessary changes.

“We focus on solving problems in the natural language environment instead of the formal environment. This eliminates the need to implement tasks, which require a lot of effort and expert knowledge in each task,” the researchers wrote.

Mind Evolution also uses the “island” method to ensure that it searches for different solutions. At each stage, the algorithm creates different sets of solutions that change in themselves. It then “migrates” the best solutions from group to group to combine and create new ones.

Mind Evolution in career planning

The researchers tested Mind Evolution against a baseline such as 1-pass, where the model generates only one response; Best-of-N, where the model generates several solutions and chooses the best one; and Sequential Revisions +, a method of revision where 10 answers are given independently, then they are reviewed separately for 80 revisions. Sequential Revisions + is the closest to Mind Evolution, although it does not have the genetic algorithm part to combine the best parts of the answer found. For more information, they also include a basic 1-step tutorial that uses it OpenAI o1-preview.

The researchers conducted many tests quickly and cheaply Gemini 1.5 Flash. They explored two ways, of course Gemini 1.5 Pro This version is used when the Flash version cannot handle the problem. This two-tier approach offers a lower cost than using the Pro version for every problem.

The researchers tested Mind Evolution on several levels of natural language processing for tasks such as travel and meeting planning. Previous research shows that LLMs cannot successfully perform these tasks without the support of professional translators.

For example, Gemini 1.5 Flash and o1-preview achieve a success rate of 5.6% and 11.7% on TravelPlanner, a benchmark that compares the planning of a trip based on preferences and constraints expressed in natural language. Even using Best-of-N on 800 automated solutions, Gemini 1.5 Flash only achieves a success rate of 55.6% on TravelPlanner.

In all of their tests, Mind Evolution beat the basics by a margin, especially as the tasks became more difficult.

For example, Mind Evolution achieves a success rate of 95% on TravelPlanner. On the Trip Planning scale, which involves making a trip of the cities to be visited with a number of days in each category, Mind Evolution scored 94.1% in the test while other methods reached a success rate of 77%. Interestingly, the difference between Mind Evolution and other methods increases as the number of cities grows, showing that they are able to handle complex planning tasks. With these two methods, Mind Evolution reached the average score in all benchmarks.

Mind Evolution also proved an inexpensive way to solve natural language processing problems, using a fraction of the symbols used by Sequential-Revision +, the only method that comes close to its performance.

“Overall, these results show the clear advantage of a revolutionary approach that combines extensive research, through passive research, and deep research that enables LLM to develop solutions,” the researchers wrote.