An AI system to help scientists write expert-level empirical software (arxiv.org)

Researchers have developed an AI system that automates the creation of expert-level empirical software to accelerate scientific discovery, addressing the traditional bottleneck of slow, manual coding for computational experiments. Leveraging a Large Language Model (LLM) combined with a Tree Search (TS) algorithm, the system systematically optimizes a defined quality metric, efficiently exploring vast solution spaces to produce high-quality, innovative software tailored to specific scientific problems. This approach has demonstrated strong performance across diverse domains, generating 40 novel single-cell data analysis methods in bioinformatics that outperformed top human-developed benchmarks. In epidemiology, it devised 14 forecasting models that surpassed CDC ensemble predictions for COVID-19 hospitalizations. Beyond these, the system achieved state-of-the-art results in geospatial analysis, neural activity prediction in zebrafish, time series forecasting, and numerical integration. By autonomously integrating complex research ideas and creating novel solutions, this AI system represents a significant advancement toward automating scientific software development, potentially accelerating research cycles and expanding the reach of computational tools across multiple scientific disciplines.