Project collaborators
- Shawn Conley, Tatiane Severo Silva, and Spyridon Mourtzinis, Department of Plant and Agroecosystem Sciences
- Jason Lo, Data Science Institute
- Damon Smith, Department of Plant Pathology
Project start and end dates
February 2025-May 2025
Project summary
“Jason Lo with the DSI played a key role in crafting a robust soybean farm management plan, expertly leveraging LLMs with meta-analysis to ensure accuracy and relevance. His precision and thoughtful approach elevated every phase of the project.”
—Shawn Conley, Department of Plant and Agroecosystem Sciences, UW–Madison
Boosting food production without expanding farmland is a key to sustainability. To help farmers apply the latest research, we tested whether AI tools like large language models (LLMs) can turn research studies into practical advice. Using U.S. soybean farming as a case, we built a human-in-the-loop system that combines AI with expert review to ensure grounded, reliable results.
To develop an LLM-assisted soybean farm management plan, we conducted a systematic literature review following established meta-analysis protocols. Using the PICO framework (Population, Intervention, Comparator, Outcome), we searched the Web of Science Core Collection for U.S.-based studies published between 2015 and 2025, focused on soybean yield responses to management practices. The search was refined iteratively for accuracy and filtered for English-language, field-based primary research.
Four LLMs (GPT-4.1-mini, Gemini 2.5, Deepseek, LLaMA3.3) independently screened studies against six inclusion criteria, with expert arbitration resolving disagreements and validating results on a 20% sample. Experts defined 10 key farm management questions, each decomposed into sub-questions via in-context learning. Two top-performing LLMs (OpenAI, Gemini) extracted answers, supporting evidence, and relevance assessments.
Papers identified by both models as relevant were included. Extracted outputs were passed through an inconsistency detection step, with expert review ensuring accuracy. All validated insights from this process were then compiled and fed into the LLM to generate a final, comprehensive soybean farm management plan. This human-in-the-loop pipeline ensured the plan was grounded in evidence, relevant, and practically useful.
Our system screened research accurately, though Gemini’s Deep Research tool produced stronger general recommendations. This approach shows promise but highlights the need for AI systems that earn farmer trust and deliver field-ready guidance.
Project deliverables
- Mourtzinis, S., Silva, T.S., Lo, J.C.M. et al. A human in the loop approach to applying large language models for farm management insight. Sci Rep (2025). https://doi.org/10.1038/s41598-025-30991-6