Skip to content
#AI

AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms (AI column)

#AI #Software #Engineering #Architecture #Agents #Math #ML

In mid-May, Google released story About his agent AlphaEvolve, which is able to develop advanced algorithms for solving optimization problems. In fact, it’s an evolving coding agent based on Google’s LLM pair. A more detailed account of the insides of this agent is available in a separate scientific (on 44 page). This AlphaEvolve algorithm is the result of a collaborative effort by the Google DeepMind team led by eight leading researchers who contributed equally: Alexander Novikov, Ngân Vй, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov and Borislav Kozlovskii. Other authors include Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli and Matej Balog. This team brought together specialists in the field of machine learning, algorithmic optimization and mathematical modeling.

If we talk about the scheme of the agent, the key points are as follows: 1. The system uses two types of models. Gemini Flash is a faster and more efficient low latency model that maximizes the amount of ideas generated per unit of time. Gemini Pro – a more powerful model, providing periodic high-quality offers for breakthrough solutions 2. The system uses an evolutionary approach to code modification. AlphaEvolve uses an evolutionary framework where agents propose programs that implement algorithmic solutions in the form of code. These solutions are automatically evaluated using something like fitness functions. 3.Automated evaluation The system includes automated appraisers who check, run and evaluate proposed programs against objective metrics. This quantifies the accuracy and quality of each solution, making the system particularly effective in areas where progress can be clearly and systematically measured. Here is a quote from the scientific article I cited above.

The LLM-directed evolution process is grounded using code execution and automatic evaluation

In the article, the guys talk about the cool results of the model

  • Improving the efficiency of data centersAlphaEvolve has developed a heuristic for the Borg orchestrator that continuously restores the average. 0,7Percentage of Google’s global computing resources. - Chip optimization: The system proposed to rewrite the Verilog code to remove excess bits in the arithmetic matrix multiplication scheme, which was integrated into the new version of the TPU.
  • Accelerating learning of LLM modelsAlphaEvolve Accelerates Key Component of Gemini Architecture 23%, which led to a reduction in Gemini training time 1%
  • New algorithm for multiplication of matricesAlphaEvolve discovers algorithm for multiplication of complex matrices 4×4 use-all 48 scalar multiplications that exceed Strassen's algorithm 1969 The year that was the top one before.

These achievements are based on previous research

  • AI Co-scientist (2025) Multi-agent system based on Gemini 2.0It is designed to help scientists generate new hypotheses and research proposals.
  • AlphaTensor (2022) A system from DeepMind that specializes in automatically detecting new matrix multiplication algorithms using deep reinforcement learning.
  • AlphaZero (2018) The deep reinforcement learning technology underlying these systems originates from AlphaZero, the DeepMind self-learning system that mastered board games such as go, chess and shogi.

As a result, the authors of this agency system have big plans for its further development, but in the current state the results look very cool and are applicable to a wide range of tasks. In fact, the authors showed how to move from specialized systems to more flexible agents.

#AI #Software #Engineering #Architecture #Agents #Math #Software #ML