AlphaEvolve orchestrates LLMs (Large Language Models) using Evolutionary algorithm to automatically modify and improve code as an autonomous coding agent.
- Humans provide the problem, evaluation metrics, and initial code
- Generate code patches (diffs) from LLMs
- Execute, evaluate, and store superior results in database
- Iteratively evolve considering both optimality and diversity
Results are impressive
- Mathematics & Algorithms: Discovered new algorithm for 4×4 complex matrix multiplication using 48 multiplications (previously 49) after 56 years.
- Scientific Problems: Found new optimal/improved solutions in about 20% of 50+ open problems (e.g., 11-dimensional kissing number 593, improved upper bound for Erdős minimum overlap problem).
- Industrial Applications
- Improved Google datacenter scheduler heuristics → 0.7% resource recovery across company.
- Automated kernel tiling heuristics design for Gemini(LLM) training → 23% faster kernels, 1% reduction in total training time.
- Achieved tangible gains in hardware/compiler optimization including TPU circuits and XLA IR.
Limitation
Areas that require human experimentation (where automatic evaluation is not possible) are still challenging, and for large problem sizes, there are memory and computational limitations that need further optimization.
While this is a result of recursive diverse calls by the AI Agent, it's unclear whether the evolutionary algorithm is truly necessary or meaningful.
example ‣