AlphaEvolve: A Gemini-based coding agent expanding its impact across multiple fields

AlphaEvolve is a Gemini-based coding agent that started with advanced algorithm design and has since widened its scope to open problems in math and computer science, Google's internal infrastructure, and industrial applications.

In genomics, it was used to improve DeepConsensus, reducing variant-calling errors by 30%, and in the power grid's AC Optimal Power Flow problem, it raised the rate of finding feasible solutions from 14% to over 88%. In earth science, it automated Earth AI optimization, improving overall accuracy of natural disaster risk prediction across 20 categories—including wildfires, floods, and tornadoes—by 5%.

Results continued at the research frontier as well.

• Proposed circuits for the Willow quantum processor with 10x lower error than the existing baseline
• Contributed to solving Erdős problems together with Terence Tao
• Used to improve lower bounds for the Traveling Salesman Problem, Ramsey Numbers, and more

Within Google's infrastructure, it was applied to TPU design, cache replacement policies, LSM-tree compaction in Spanner, and compiler optimization. Notably, the cache replacement policy work, which previously took months, was completed in just two days, and the Spanner-side improvement reduced write amplification by 20%.

Commercial adoption also expanded.

• Klarna: 2x improvement in training speed
• FM Logistic: 10.4% improvement in route efficiency, saving over 15,000km annually
• Schrödinger: about 4x speedup in both MLFF training and inference

Google DeepMind stated that it plans to further expand AlphaEvolve's capabilities and apply it broadly to external challenges.