Google AI aims to make best-in-class scientific software even better

Nature NewsCenterEN 4 min read 100% complete by Matthew HutsonOctober 25, 2025 at 04:27 PM

AI Summary

long article 4 min

Google researchers have developed an AI-guided software-evolution system to enhance scientific software performance. The team created evolutionary 'trees' of software tools for six tasks, with each node representing a program assessed by standard benchmarks. By prompting large language models (LLMs) to improve existing programs and providing them with research summaries and specialist knowledge, the researchers generated new nodes that outperformed state-of-the-art tools in several cases. The workflow aims to automate scientific coding processes, which are typically time-consuming and technical. Google is working on making this system accessible to scientists for broader application.

Key Topics & Entities

Nature 50% University of British Columbia 80% Google ×4100%

Keywords

artificial intelligence 100% software evolution 90% scientific software 80% google ai 70% large language model (llm) 60% performance improvement 50% code optimization 40%