STELLA: Self-Evolving LLM Agent for Biomedical Research

TL;DR

STELLA is introduced, a self-evolving AI agent designed to overcome limitations in AI agents that can learn and grow, dynamically scaling their expertise to accelerate the pace of biomedical discovery.

Abstract

The rapid growth of biomedical data, tools, and literature has created a fragmented research landscape that outpaces human expertise. While AI agents offer a solution, they typically rely on static, manually curated toolsets, limiting their ability to adapt and scale. Here, we introduce STELLA, a self-evolving AI agent designed to overcome these limitations. STELLA employs a multi-agent architecture that autonomously improves its own capabilities through two core mechanisms: an evolving Template Library for reasoning strategies and a dynamic Tool Ocean that expands as a Tool Creation Agent automatically discovers and integrates new bioinformatics tools. This allows STELLA to learn from experience. We demonstrate that STELLA achieves state-of-the-art accuracy on a suite of biomedical benchmarks, scoring approximately 26\% on Humanity's Last Exam: Biomedicine, 54\% on LAB-Bench: DBQA, and 63\% on LAB-Bench: LitQA, outperforming leading models by up to 6 percentage points. More importantly, we show that its performance systematically improves with experience; for instance, its accuracy on the Humanity's Last Exam benchmark almost doubles with increased trials. STELLA represents a significant advance towards AI Agent systems that can learn and grow, dynamically scaling their expertise to accelerate the pace of biomedical discovery.

Figures (2)

• Figure 1: Overall Framework of STELLA, a self-evolving LLM Agent for Biomedical Research. (A) STELLA leverages four key agents, including Manager Agent, Dev Agent, Critic Agent, and Tool Creation Agent. The manager agent coordinates all agents and curates a reasoning template library to leverage successful reasoning experience; dev agent focuses on environment building, code creation, model training, and report writing; critic agents reflects on the intermediary results and provide suggestions; tool creation agent identifies the gap of agent capabilities and create new tools stored in Tool Ocean. Human expert and wet experiment results can provide valuable feedback and guidance in the loop. (B and C) Two key features of STELLA's self-evolving mechanisms. The Template Library evolves by including successful previous examples; the Tool Ocean evolves from simple predefined tools during agent inference.
• Figure 2: (A) Benchmark results of Stella with state-of-the-art LLMs and agents on Humanity's Last Exam: Biomedicine and LAB-Bench: DBQA and LitQA. (B) Test-time self-evolving effects on Benchmarks. The computation budget indicates the number of trials. The reported results represent the average accuracy across three independent evaluation runs.