Scientific Discoveries by LLM Agents

Rachel So

Keywords: big science

TL;DR: A survey on how LLM-powered AI agents now autonomously conduct full scientific research pipelines to make verifiable discoveries in fields like chemistry, biology, and materials science.

Abstract: Large Language Models (LLMs) have evolved from text generators into sophisticated autonomous agents capable of conducting independent scientific research. This paper reviews the current landscape of LLM-driven scientific discovery, where AI agents can now execute the entire research pipeline, including reading scientific literature, forming novel hypotheses, designing experiments, interfacing with laboratory tools and simulators, analyzing data, and interpreting results. A key advancement is the deployment of multi-agent systems, where specialized agents collaborate in roles such as 'scientist,' 'critic,' and 'evaluator' to tackle complex challenges beyond the scope of individual agents. We survey domain-specific applications and highlight validated discoveries, including the autonomous synthesis of novel chemical compounds and materials, the design of functional nanobodies for SARS-CoV-2 variants, and the automation of complex bioinformatics analyses. The development of end-to-end research systems that can progress from an initial idea to a full, peer-reviewed publication demonstrates a paradigm shift in the automation of science. Despite these successes, significant challenges remain, including performance degradation on highly complex causal reasoning tasks. Future directions point toward creating more robust, causally-aware agents and enhancing human-AI collaboration to accelerate scientific breakthroughs.

Submission Number: 49