AlphaEvolve continuously refines its output by scoring each of Gemini’s suggestions, discarding weaker attempts and iteratively improving the stronger ones.
Ahead of the Google I/O 2025 annual developer conference, starting May 20, Google has announced its new AI agent called AlphaEvolve. The company introduces the AI agent as “an evolutionary coding agent powered by large language models for general-purpose algorithm discovery and optimisation.”
It explains that AlphaEvolve combines the inventive problem-solving strengths of Google’s Gemini models with automated evaluators that validate solutions, employing an evolutionary system to refine and build on the most promising concepts.
While LLMs are often hit-or-miss when it comes to generating code, AlphaEvolve takes a different approach. It continuously refines its output by scoring each of Gemini’s suggestions, discarding weaker attempts and iteratively improving the stronger ones. This evolutionary process enables the system to produce highly optimised algorithms, many of which outperform the best human-written alternatives in terms of speed or accuracy.
ALSO READ: Why Are F1 Teams Turning To Big Data Analytics?
One standout example of AlphaEvolve’s capabilities, as shared by the company, is its role in improving Google’s job scheduling software, which allocates computing tasks across millions of servers worldwide. According to DeepMind, the refined algorithm has been running in production across Google’s global data centres for over a year, unlocking a 0.7% gain in computing efficiency—a modest-sounding figure, but a massive boost at Google’s scale.
AlphaEvolve on Cutting Down AI Hallucinations
AlphaEvolve also addresses one of the major pitfalls of modern AI: hallucinations. Most AI systems, due to their probabilistic nature, sometimes fabricate confident but false answers. In fact, newer models, including OpenAI’s o3, have demonstrated an increased tendency to do so.
To combat this, AlphaEvolve introduces an automated evaluation layer. It prompts the model to generate multiple potential answers, then critiques and scores them based on accuracy, effectively filtering out unreliable responses.
Google DeepMind in its blogpost stated, “AlphaEvolve verifies, runs and scores the proposed programs using automated evaluation metrics. These metrics provide an objective, quantifiable assessment of each solution’s accuracy and quality. This makes AlphaEvolve particularly helpful in a broad range of domains where progress can be clearly and systematically measured, like in math and computer science.”
ALSO READ: Stanford Open-Sources Controllable Generative Language AI
