Google DeepMind has unveiled its latest breakthrough in artificial intelligence: AlphaEvolve, an advanced AI agent that not only tackles unsolved theoretical problems in mathematics and computer science but is already improving critical real-world operations inside Google.
Built on the Gemini 2.0 family of large language models (LLMs), AlphaEvolve represents a new step in AI-driven scientific discovery. Unlike typical LLMs that often generate hit-or-miss code suggestions, AlphaEvolve uses a tightly controlled iterative process. It generates multiple candidate solutions, rigorously scores them for efficiency and accuracy, tweaks the most promising ones, and repeats the cycle until it produces an optimal algorithm.
“You can think of it as a super coding agent,” says Pushmeet Kohli, VP at Google DeepMind and head of its AI for Science division. “It doesn’t just suggest code—it produces solutions that even human experts may have never conceived.”
Real-World Impact
One of AlphaEvolve’s most immediate applications is already live inside Google’s infrastructure. Over the past year, Google has been using an AlphaEvolve-designed algorithm to optimize job scheduling across its global network of data centers. The result: a 0.7% increase in resource efficiency. While that figure may seem modest, at Google’s operational scale it represents enormous computational savings.
AlphaEvolve also delivered a new power management algorithm that reduces energy consumption in Google’s Tensor Processing Unit (TPU) chips, and even helped speed up the training process for Google’s flagship Gemini models by optimizing specific computational tasks involved in LLM training.
A Legacy of AI-Led Discovery
AlphaEvolve follows a series of DeepMind innovations in algorithm discovery. In 2022, AlphaTensor broke a 50-year-old record by finding a faster way to multiply matrices—a cornerstone operation in computer science and AI. In 2023, AlphaDev uncovered faster ways to execute low-level computer instructions used billions of times per second across global computing systems.
However, AlphaEvolve takes things a step further by being a general-purpose problem solver. It can be applied to virtually any task where solutions can be written in code and evaluated computationally.
The Evolutionary Process Behind AlphaEvolve
Here’s how AlphaEvolve works: A user feeds it a problem description, with optional hints like previous algorithms or example solutions. AlphaEvolve then uses Gemini 2.0 Flash—Google’s smallest and fastest LLM—to generate thousands of potential code solutions. Each one is executed and scored against pre-defined benchmarks: Does it produce correct results? Does it run faster? Is it more efficient?
The best-performing solutions from each round are fed back into Gemini for further refinement. To prevent the AI from getting stuck in dead-ends, AlphaEvolve occasionally reinserts previous candidates into the pool. When Gemini Flash hits a wall, the system escalates to Gemini 2.0 Pro, the most powerful LLM in Google’s lineup, to unlock more complex solutions.
This evolutionary process continues until no further improvements emerge, often resulting in algorithms that surpass human-written benchmarks.
Outperforming in Math and Science
The research team tested AlphaEvolve across more than 50 mathematical challenges, including:
Matrix multiplication: AlphaEvolve not only matched but beat AlphaTensor’s world-record solution for multiplying 4×4 matrices, producing results that work across a wider range of numbers—not just 0s and 1s.
Fourier analysis: Crucial for data compression technologies like video streaming.
The Minimum Overlap Problem: An unsolved number theory puzzle first posed by legendary mathematician Paul Erdős in 1955.
Kissing numbers: A geometry problem with implications in materials science, chemistry, and cryptography.
In total, AlphaEvolve matched the best known human solutions in 75% of cases and outperformed them in 20%, delivering faster or more efficient algorithms.
A Broader Shift in Research
While AlphaEvolve’s results are impressive, researchers note that the tool offers limited theoretical insight into how or why its solutions work—something still crucial for advancing fundamental human understanding of math and science.
“It’s a bit of a black box,” says Manuel Kauers, a mathematician at Johannes Kepler University, Austria, whose team recently produced similar matrix multiplication results using different methods. “But progress is progress.”
The technology is also currently limited to problems where solutions can be quantitatively evaluated by a machine, making it less useful for subjective or interpretive tasks.
Still, Google DeepMind researchers believe AlphaEvolve represents a paradigm shift in how future algorithms—and possibly scientific theories—will be discovered.
“We’re not done,” says Kohli. “There’s much further to go in seeing how powerful this kind of approach can become.”
Prepared by Navruzakhon Burieva
