Skip to content
back to the archive page
#AI

AlphaEvolve: A Gemini-Powered Coding Agent for Designing Advanced Algorithms (Category AI)

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

In mid-May, Google published an interesting article about AlphaEvolve, its agent for developing advanced algorithms for optimization problems. In effect, it is an evolving coding agent built around a pair of Google LLMs. A separate 44-page research paper explains its internals in more detail. AlphaEvolve is a collective effort by Google DeepMind, led by eight researchers with equal contributions: Alexander Novikov, Ngân Vũ, 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. The team brings together expertise in machine learning, algorithmic optimization, and mathematical modeling.

The key aspects of the agent’s design are:

1. Two types of models

  • Gemini Flash: a faster, efficient, low-latency model that maximizes the number of ideas generated per unit of time.
  • Gemini Pro: a more powerful model that periodically contributes high-quality suggestions for breakthrough solutions.

2. An evolutionary approach to modifying code AlphaEvolve uses an evolutionary framework in which agents propose programs implementing algorithmic solutions. Those solutions are automatically assessed using something like fitness functions.

3. Automated evaluation Automated evaluators check, run, and score the proposed programs against objective metrics. This quantifies each solution’s accuracy and quality, making the system particularly effective where progress can be measured clearly and systematically. As the paper puts it:

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

The authors report some impressive results:

  • Data-center efficiency: AlphaEvolve developed a heuristic for the Borg orchestrator that continually recovers an average of 0.7% of Google’s worldwide computing resources.
  • Chip optimization: it proposed rewriting Verilog code to remove redundant bits in a matrix-multiplication arithmetic circuit. This was incorporated into a new TPU version.
  • Faster LLM training: it accelerated a key Gemini architecture component by 23%, reducing Gemini’s training time by 1%.
  • A new matrix-multiplication algorithm: it found a way to multiply complex 4×4 matrices with just 48 scalar multiplications, improving on Strassen’s 1969 algorithm, the previous leading result.

These achievements build on earlier research:

  • AI Co-scientist (2025): a multi-agent system based on Gemini 2.0, designed to help scientists generate hypotheses and research proposals.
  • AlphaTensor (2022): DeepMind’s system for discovering new matrix-multiplication algorithms through deep reinforcement learning.
  • AlphaZero (2018): the deep reinforcement-learning technology behind these systems traces back to DeepMind’s self-learning system that mastered board games including Go, chess, and shogi.

The authors have ambitious plans for this agent system. Even now, its results look impressive and applicable to many problems. In effect, they have demonstrated a path from specialized systems to more flexible agents.

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