English

NewsGoogleGeminiAlphaEvolve

Google Reveals Multi-Field Achievements of Gemini-Powered Coding Agent AlphaEvolve

This article is a translation. Read the Japanese original

Google has released the latest results for "AlphaEvolve," a Gemini-powered coding agent. The agent is a system designed to support the design and optimization of algorithms. The published cases demonstrate that it has brought concrete progress across multiple scientific and business domains.

In the field of genomics, AlphaEvolve was utilized to improve "DeepConsensus," a DNA sequence error correction model. It was reported that this led to a 30% reduction in variant detection errors. Aaron Wenger, Senior Director at PacBio, stated that this improvement in accuracy contributes to the discovery of hidden disease-causing mutations.

Regarding power grid optimization, when applied to the AC Optimal Power Flow problem, the discovery rate of feasible solutions using GNN models improved from 14% to over 88%. In the field of Earth sciences, the accuracy of natural disaster risk prediction reportedly improved by 5%, a figure integrating prediction accuracy across 20 categories.

In quantum physics, the agent proposed quantum circuits that enable the execution of complex molecular simulations on the Willow quantum processor. Google stated that errors were 10 times lower compared to conventional benchmarks. Additionally, in collaboration with mathematicians such as Terence Tao, the agent contributed to solving Erdős problems and breaking records for challenges such as the Traveling Salesperson Problem.


Source: AlphaEvolve: Gemini-powered coding agent scaling impact across fields (HN 327pt, 149 comments) (HN Search (backfill), 2026-05-08)