Elamur Research⌕
☰
home / papers / 2025 / STELLA: Self-Evolving LLM Agent for Biomedical Research

STELLA: Self-Evolving LLM Agent for Biomedical Research

Paper · 2025 · Ruofan Jin, Zaixi Zhang, Mengdi Wang et al.

Summary

Biomedical research relies on a large and constantly changing collection of databases, software packages and publications. Most current LLM agents work with a fixed, hand-picked toolset, so they struggle to keep up. STELLA is a multi-agent system built to improve itself through use. A Manager Agent plans a stepwise reasoning pathway. A Dev Agent writes and runs Python code in its own conda environments. A Critic Agent reviews intermediate results and suggests fixes. A Tool Creation Agent closes capability gaps by locating, testing and adding new tools.

Two resources grow as the system works. The Template Library stores condensed workflows from solved problems. The Tool Ocean collects executable capabilities. On sampled subsets of Humanity's Last Exam (Biomedicine) and LAB-Bench (DBQA, LitQA), STELLA scores higher than the LLMs and agents it is compared with. Its HLE accuracy roughly doubles as the number of trials increases. The authors argue that agents which build up tools and strategies over time can scale with the field better than static ones.

Key points

  • Reported accuracy is about 26% on HLE: Biomedicine, about 54% on LAB-Bench DBQA and about 63% on LitQA. The margin over the next-best system is stated as up to 6 points in the abstract but up to 8 points in the main text.
  • Evidence for self-evolution: HLE accuracy rises from 14% to 26% as the trial budget grows, averaged over three independent runs.
  • The evaluation follows the Biomni and OriGene setups. For LAB-Bench, the authors used the same 12.5% sample of DBQA and LitQA, with multiple-choice answers and an abstention option. For HLE, they used 50 questions covering 14 biology and medicine subfields.
  • The Tool Ocean contains three types of tools:
    • Database queries (PubMed, ClinVar, PDB).
    • Foundation-model interfaces (AlphaFold 3, scGPT, ESM3).
    • Custom scripts, for example for network analysis and data integration.
  • In a case study on cisplatin resistance, STELLA analysed scRNA-seq data from before treatment and after relapse. It then built a virtual perturbation screen based on the STATE cell model and identified the transcription factor MTF1 as the key regulator.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Decision Sciences · 306 worksBiochemistry, Genetics and Molecular Biology · 273 worksMolecular Biology · 270 worksMedicine · 288 worksHealth Informatics · 227 worksInformation Systems and Management · 96 works

Cite

Cite as Ruofan Jin, Zaixi Zhang, Mengdi Wang and Le Cong. 2025. STELLA: Self-Evolving LLM Agent for Biomedical Research. arXiv:2507.02004.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2507.02004 · arXiv 2507.02004

@misc{jin2025stella,
  title = {{STELLA: Self-Evolving LLM Agent for Biomedical Research}},
  author = {Ruofan Jin, Zaixi Zhang, Mengdi Wang and Le Cong},
  year = {2025},
  doi = {10.48550/arxiv.2507.02004},
  eprint = {2507.02004},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2507.02004}
}

References

23 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. Botvinik-Nezer, et al., Variability in the analysis of a single neuroimaging dataset by many teams. Nature 582 (7810), 84–88 (2020). find
  2. I. Thiele, B. Ø. Palsson, A protocol for generating a high-quality genome-scale metabolic reconstruction. Nature protocols 5 (1), 93–121 (2010). find
  3. E. Gibney, R. Van Noorden, Scientists losing data at a rapid rate. Nature (2013), doi:10.1038/ nature.2013.14416. find
  4. H. Wang, et al., Scientific discovery in the age of artificial intelligence. Nature 620 (7972), 47–60 (2023). find
  5. G. Tom, S. Li, J. M. Gregoire, A. Aspuru-Guzik, Self-driving laboratories for chemistry and materials science. Chemical Reviews 124 (16), 9633–9732 (2024). find
  6. C. Peng, et al., A study of generative large language model for medical research and healthcare. NPJ digital medicine 6 (1), 210 (2023). find
  7. Y. Qu, et al., Crispr-gpt: An llm agent for automated design of gene-editing experiments. bioRxiv pp. 2024–04 (2024). find
  8. K. Swanson, W. Wu, N. L. Bulaong, J. E. Pak, J. Zou, The virtual lab: Ai agents design new sars-cov-2 nanobodies with experimental validation. bioRxiv pp. 2024–11 (2024). find
  9. Y. Roohani, et al., Biodiscoveryagent: An ai agent for designing genetic perturbation experiments. International Conference on Learning Representations (2025). \#\# 10. E. Wang, et al., Txgemma: Efficient and agentic llms for therapeutics. arXiv preprint arXiv:2504.06196 (2025).
  10. S. Gao, et al., Txagent: An ai agent for therapeutic reasoning across a universe of tools. arXiv preprint arXiv:2503.10970 (2025).
  11. Y. Xiao, et al., Cellagent: An llm-driven multi-agent framework for automated single-cell data analysis. BioRxiv pp. 2024–05 (2024). find
  12. Z. Zhang, et al., OriGene: A Self-Evolving Virtual Disease Biologist Automating Therapeutic Target Discovery. bioRxiv pp. 2025–06 (2025). \#\# 14. K. Huang, et al., Biomni: A General-Purpose Biomedical AI Agent. bioRxiv pp. 2025–05 (2025). find
  13. J. Qiu, et al., Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution. arXiv preprint arXiv:2505.20286 (2025). — [arXiv:2505.20286](https://arxiv.org/abs/2505.20286) find
  14. J. Wei, et al., Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, 24824–24837 (2022).
  15. S. Yao, et al., React: Synergizing reasoning and acting in language models, in International Conference on Learning Representations (ICLR) (2023).
  16. X. Wang, et al., Executable code actions elicit better llm agents, in Forty-first International Conference on Machine Learning (2024). \#\# 19. L. Phan, et al., Humanity’s last exam. arXiv preprint arXiv:2501.14249 (2025). — Humanity's Last Exam
  17. J. M. Laurent, et al., Lab-bench: Measuring capabilities of language models for biology research. arXiv preprint arXiv:2407.10362 (2024).
  18. A. K. Adduri, et al., Predicting cellular responses to perturbation across diverse contexts with STATE. bioRxiv (2025), doi:10.1101/2025.06.26.661135v1, https://www.biorxiv.org/content/10.1101/2025.06.26.661135v1. \#\# 22. J. White, PubMed 2.0. Medical reference services quarterly 39 (4), 382–387 (202… — [doi:10.1101/2025.06.26.661135v1](https://doi.org/10.1101/2025.06.26.661135v1) ↗
  19. M. J. Landrum, et al., ClinVar: public archive of interpretations of clinically relevant variants. Nucleic acids research 44 (D1), D862–D868 (2016). find
  20. J. L. Sussman, et al., Protein Data Bank (PDB): database of three-dimensional structural information of biological macromolecules. Biological Crystallography 54 (6), 1078–1084 (1998). find
  21. J. Abramson, et al., Accurate structure prediction of biomolecular interactions with AlphaFold \#\# 3. Nature 630 (8016), 493–500 (2024). find
  22. H. Cui, et al., scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods 21 (8), 1470–1480 (2024). find
  23. T. Hayes, et al., Simulating 500 million years of evolution with a language model. Science p. eads0018 (2025). \#\# 28. L. Phan, et al., Humanity’s last exam. arXiv preprint arXiv:2501.14249 (2025). \#\# 29. Google, Gemini Pro, Web page (2024), — Humanity's Last Exam https://deepmind.google/models/gemini/pro/, accessed: July… ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/jin-2025-stella-self-evolving-llm-agent/