AI System Creates Expert-Level Scientific Software
An AI system named Empirical Research Assistance (ERA) has been created by researchers to enhance computational experiments by generating expert-level scientific software. By integrating a Large Language Model (LLM) with Tree Search (TS), ERA systematically optimizes quality metrics and explores solution spaces. In the field of bioinformatics, ERA identified 40 new methods for single-cell data analysis, which exceeded the performance of leading human-created methods on a public leaderboard. Additionally, in epidemiology, ERA generated 14 models that outperformed those developed by the CDC.
Key facts
- ERA uses an LLM and Tree Search to create scientific software.
- The system aims to maximize a quality metric.
- ERA discovered 40 novel methods for single-cell data analysis.
- These methods outperformed top human-developed methods on a public leaderboard.
- ERA generated 14 models that outperformed CDC models in epidemiology.
- The system integrates complex research ideas from external sources.
- Tree Search effectiveness was demonstrated across diverse tasks.
- The research is described in arXiv:2509.06503.
Entities
Institutions
- CDC