A Penn State-led team says it used three AI agents to collect more than 4,000 measurements from food-science papers and uncover a reported relationship between two properties that shape microwave heating. The measurements were scattered across tables, figures and equations, making them hard to compare by hand.
AgentXtract Builds a Food Dataset and Identifies a Reported Relationship
The workflow, called AgentXtract, assigns agents to interpret papers, extract numerical values, and standardize and assemble the records. The study reports more than 4,000 records across more than 140 food materials in about 30 minutes, at an API cost of about $9. The paper became available online on Sept. 12, ahead of its scheduled appearance in the March 2027 issue of the Journal of Food Engineering.
Dielectric constant and loss factor are measurements of how food interacts with electromagnetic energy such as microwaves. The paper calls the link it found between them a power law, a mathematical description of how two quantities scale together. In plain terms, the compiled measurements showed the two properties changing together in a consistent pattern. The authors found that pattern in data covering more than 100 foods; it describes this dataset, not a rule already shown to hold for every food or measurement condition.
Extraction Tests Compare Models; a Prototype Retrieves Source-Linked Data
The authors compared 10 large language models from OpenAI, Google and Anthropic with a reference dataset checked by experts. All ten models had at least 95% precision, while recall varied from 79% to 99%. In this test, precision measures how often extracted values matched the reference; recall measures how much of the reference data a model recovered. The results suggest the models returned mostly supported values, but some missed a larger share of measurements. Human expertise remained part of the process: researchers identified suitable papers and checked data quality.
The team also connected the database to a prototype retrieval-augmented generation assistant. Asked about a food’s properties at a particular frequency, it can search the database and return measurements with references. This prototype uses a curated, source-linked dataset; the study does not show that a general-purpose chatbot can independently recover the same measurements from the full literature.
Follow-Up Tests and Published Code
Independent measurements would show how widely the reported pattern applies beyond the assembled dataset. The team has published AgentXtract’s code under an MIT license, and the paper says the selected papers and extracted database are available in its supplementary information.
Image: “Open microwave oven” by MichalPL, via Wikimedia Commons, CC BY-SA 4.0. Used without changes; illustrative and not part of the study.
