STELLA

A Self-Evolving Agentic System for Biomedical Discovery

Ruofan Jin1 Mingyang Xu2 Jiangnan Zhao3,4,5 Fei Meng6,7 Shiyao Chen1 Guancheng Wan8 Qingran Cai9 Yiyang He1 Yize Jiang10 Jin Han11 Yuanyuan Chen12 Wanqing Lu12 Mengyang Wang13 Zhiqian Lan14 Yuxuan Jiang14 Xiang Sheng17,18 Le Cong15,✉ Junhong Liu10,✉ Dongyao Wang6,7,✉ Kai Wang3,4,5,✉ Zaixi Zhang16,✉

1Zhejiang University 2University of Michigan 3Department of Respiratory and Critical Care Medicine, the Fourth Affiliated Hospital of School of Medicine, and International School of Medicine, International Institutes of Medicine, Zhejiang University 4Zhejiang Key Laboratory of Precision Diagnosis and Treatment for Lung Cancer 5Zhejiang-Sweden Joint Laboratory on Tumor Immunology 6The First Affiliated Hospital, University of Science and Technology of China 7National Key Laboratory of Immune Response and Immunotherapy 8University of California, Los Angeles 9Shanghai Jiao Tong University 10Microcyto 11Nanjing University 12Tianjin University of Science and Technology 13Peking University 14The University of Hong Kong 15Stanford University 16The Hong Kong University of Science and Technology 17University of Chinese Academy of Sciences 18Key Laboratory of Engineering Biology for Low-Carbon Manufacturing, Institute of Industrial Biotechnology, Chinese Academy of Sciences

✉Corresponding authors

Paper Code Try STELLA

Online agent powered by BioClaw, built on STELLA.

A self-evolving multimodal agent system for biomedical research — it plans, executes, critiques its own work, and builds the tools it is missing.

60+
tools in the Tool Ocean
117
tasks in the Tool Creation Benchmark
4
cell lines validating BTN3A1 by CRISPR
17%→82%
robotic lab automation success rate
01Abstract

Most AI agents are limited to a fixed toolbox. STELLA is a self-evolving multimodal agent system for biomedical research: new tools join a growing Tool Ocean, and successful runs become reusable workflow templates — so it gets more capable the more it is used.

STELLA achieves leading performance on the Tool Creation Benchmark, LabBench and HLE, and its findings hold up in the lab: it identified BTN3A1 as a regulator of NK cell function in acute myeloid leukemia (CRISPR-validated in four cell lines), produced enzyme variants with over two-fold higher catalytic activity, and raised robotic lab automation success from 17% to 82%.

02Method

A team of agents that grows its own toolkit

STELLA organizes research as a loop between coordinated agents and an evolving library of tools and workflows. Every run can leave the system more capable than before.

Research goal
Manager Agent
Decomposes goals · retrieves templates · coordinates agents
Dev Agent
Analyze · code · query
Critic Agent
Evaluate · refine
Tool Creation Agent
Build missing tools
Self-evolving layer
Tool Ocean
60+ prebuilt & self-evolved tools
Template System
Workflows from successful runs

↺ Reused in future runs

  1. Manager Agent

    Breaks a research objective into concrete steps, retrieves matching workflow templates, and coordinates the specialist agents.

  2. Dev Agent

    Carries out bioinformatic analyses, writes and runs code, and queries biological databases and the literature.

  3. Critic Agent

    Assesses the quality of intermediate and final results and proposes concrete improvements.

  4. Tool Creation Agent

    When existing tools are not enough, designs and builds new ones and adds them to the Tool Ocean.

  5. Tool Ocean

    A library of 60+ prebuilt and self-evolved tools that keeps expanding as STELLA works.

  6. Template System

    Saves workflows from successful runs and generates new templates automatically, so solved problems become reusable.

03Results

From benchmarks to the bench

Benchmarks

Leading

Tool Creation Benchmark

117 tasks

Leading

LabBench

 

Leading

HLE

Humanity's Last Exam

Validated in the lab

Target discovery

BTN3A1 regulates NK cell function in AML

Identified by STELLA and validated by CRISPR in four cell lines.

Enzyme engineering

Over 2× higher catalytic activity

Enzyme variants with more than two-fold improved catalytic activity.

Lab automation

Robotic success rate: 17% → 82%

Success rate of robotic lab automation rose from 17% to 82%.

05Citation

BibTeX

@article{jin2025stella,
  title   = {STELLA: A Self-Evolving Agentic System for Biomedical Discovery},
  author  = {Jin, Ruofan and Xu, Mingyang and Zhao, Jiangnan and Meng, Fei and Chen, Shiyao and Wan, Guancheng and Cai, Qingran and He, Yiyang and Jiang, Yize and Han, Jin and Chen, Yuanyuan and Lu, Wanqing and Wang, Mengyang and Lan, Zhiqian and Jiang, Yuxuan and Sheng, Xiang and Cong, Le and Liu, Junhong and Wang, Dongyao and Wang, Kai and Zhang, Zaixi},
  journal = {bioRxiv},
  year    = {2025},
  doi     = {10.1101/2025.07.01.662467},
  url     = {https://www.biorxiv.org/content/10.1101/2025.07.01.662467v2}
}