Eubiota/eubiota-planner-8b

🤗 HuggingFace ｜ 🌐 Website ｜ 📄 Paper ｜ 💻 GitHub ｜ 🚀 App

---

Eubiota-Planner-8B is a specialized 8-billion parameter language model fine-tuned for autonomous microbiome discovery. It serves as the core planner module in the Eubiota agentic AI framework, orchestrating multi-agent reasoning and tool-grounded scientific exploration.

Model Description

Property	Details
Model Type	Causal language model fine-tuned for agentic planning
Base Model	Qwen3-8B
Training Method	GRPO-MAS (Group Relative Policy Optimization for Multi-Agent Systems)
Language	English
Developed By	Stanford University
Affiliations	Department of Biomedical Data Science, Department of Computer Science, Department of Microbiology and Immunology, Institute for Human-Centered AI (HAI)

Intended Use

Primary Use Cases

• Microbiome Hypothesis Generation: Formulating mechanistic hypotheses about host-microbe interactions
• Experimental Design: Planning and orchestrating scientific workflows for microbiome research
• Multi-Tool Coordination: Selecting and sequencing domain-specific tools (PubMed, KEGG pathways, etc.) for complex scientific queries

Training

Training Data

The model was trained on a mixture of four domain datasets:

Dataset	Purpose	Source
Microbio-Bench	Curated microbiome reasoning	Eubiota/microbio-bench
PubMedQA	Medical-biology reasoning	qiaojin/PubMedQA
MedQA-USMLE	Medical-biology reasoning	GBaker/MedQA-USMLE-4-options
DeepMath-103K	Mathematical reasoning	zwhe99/DeepMath-103K
Natural Questions	Agentic search	RUC-NLPIR/FlashRAG_datasets

Training Method

The model is trained using GRPO-MAS (Group Relative Policy Optimization for Multi-Agent Systems), a reinforcement learning approach designed for multi-agent coordination. This method enables the planner to learn effective tool selection and sequencing strategies through outcome-driven refinement.

Integrated Tools

Eubiota-Planner-8B orchestrates 18 domain-specific tools spanning web and literature search, biological databases, laboratory resources, and computation utilities.

Category	Tools	Description
Web Search	Google Search, Perplexity Search, Wikipedia Search, URL Page Retrieval	Grounded web search with citation support
Literature Search	PubMed Search	Retrieves relevant papers from PubMed and produces citation-grounded summaries
KEGG Databases	KEGG Gene, KEGG Drug, KEGG Disease, KEGG Organism	Query tools for genes, drugs, diseases, and organisms
MDIPID Databases	MDIPID Gene View, MDIPID Disease View, MDIPID Microbe View	Microbiota-drug-disease association queries from multiple perspectives
Laboratory Tools	Transposon Screen Database, Protocol Library	Access to pre-computed gene-phenotype rankings (1,836 genes) and 55 curated microbiome/immunology experimental protocols
Document Utilities	Doc Context Search, Database Context Search	File and document context retrieval
Computation	Python Execution, LLM Summarization	Code execution and text synthesis capabilities

Benchmark Suite

Eubiota-Planner-8B is evaluated on six benchmarks across two categories to assess both broad biomedical competence and microbiome-specific mechanistic reasoning.

General Biomedical Competence

Benchmark	Description	Source
MedMCQA (Medicine)	Professional clinical knowledge and complex diagnostic reasoning from medical entrance examinations	Pal et al., 2022
WMDP-Bio	Expert-level capabilities in hazardous biological knowledge and biosecurity-relevant domains	Li et al., 2024

Microbiome-Specific Mechanistic Reasoning

Four domain-specific benchmarks derived from MDIPID emphasizing mechanistic linking among microbes, drugs, host pathways, and genes:

Benchmark	Task	Description
Drug-Microbe Impact (Drug-Imp)	Drug → Microbe	Determining the directional effects of drugs on microbial growth
Microbe-Protein Mechanism (MB-Mec)	Microbe → Protein	Linking microbial activity to host proteins
Protein Functional Comprehension (Prot-Func)	Protein → Function	Identifying the biological function of a protein
Protein-Gene Mapping (Prot-Gen)	Protein → Gene	Mapping expressed proteins to their encoding genes

Acknowledgements

We also thank the following open-source projects:

• VeRL for the excellent RL framework design.
• vLLM for fast LLM inference support.
• AgentFlow and Agent Lightning for early-stage exploration in multi-agent RL training.

Citation

@article{lu2026eubiota,
  title = {Eubiota: Modular Agentic AI for Autonomous Discovery in the Gut Microbiome},
  author = {Lu, Pan and Gao, Yifan and Peng, William G. and Zhang, Haoxiang and Zhu, Kunlun and Robinson, Elektra K. and Xu, Qixin and Kotaka, Masakazu and Zhang, Harrison G. and Li, Bingxuan and Shiver, Anthony L. and Choi, Yejin and Huang, Kerwyn Casey and Sonnenburg, Justin and Zou, James},
  journal = {bioRxiv},
  year = {2026},
  month = {feb},
  day = {27},
  doi = {10.64898/2026.02.27.708412},
  url = {https://www.biorxiv.org/content/10.64898/2026.02.27.708412v1},
  publisher = {Cold Spring Harbor Laboratory}
}