Press
Boilerplate — use verbatim
Where it fits, add: "AI AgriBench evaluates natural-language agronomy answers; it is distinct from similarly named multimodal or computer-vision benchmarks."
The name
Write "AI AgriBench" in full. Not "AI-AgriBench". Not "AgriBench" on its own in a headline or a caption.
Fact sheet
| What it is | A public consortium that independently benchmarks AI question-answering systems for agriculture and publishes the results |
| Led by | Center for Digital Agriculture, University of Illinois Urbana-Champaign |
| Origin | Spun out of the CropWizard project within AIFARMS, the National AI Institute for Food and Agriculture |
| Consortium launched | 28 January 2026 (original CDA announcement, March 2025) |
| Founding members | 12 |
| Leaderboards | Two: main (US, English, technical agronomy) and Smallholders |
| Source corpus | 40,000+ extension publications and research papers |
| Benchmark size, v0.5 | 45,172 pairs generated → 951 expert-reviewed → 416 retained |
| Smallholders data | 3,100+ real farmer–expert question-and-answer pairs |
| Expert review | 23 agronomists |
| Topic categories | 9: pests, diseases, weeds, nutrition, soils, seed hybrids, horticulture, water, weather |
| Peer-reviewed foundations | YourBench (arXiv:2504.01853); MIRAGE (NeurIPS 2025 Datasets & Benchmarks) |
| Funders to date | AIFARMS, CDA, Gates Foundation |
Headline statistics
- 94–98% — frontier model accuracy on technical agronomy questions, AI AgriBench v0.5 main track.
- 65–81% — the same models on real smallholder farmer questions, v0.5 Smallholders track.
- 500 million+ — farmers worldwide who may come to rely on AI advisory tools.
Cite the release and the track with every figure. Never present a main-track and a Smallholders figure as comparable.
Story angles
- "AI knows the textbook but not the field" — the 94–98% versus 65–81% gap.
- Industry–academic collaboration: Bayer, John Deere and Microsoft at the same table as Digital Green and an Indian agtech startup.
- AI for smallholders: a benchmark designed in Illinois that will eventually score Hindi, Odia and Amharic audio question answering.
- The consortium model as a template for responsible AI in other regulated advisory domains.
Logo files
TODO (Comms): the press kit does not exist yet. Produce and host here — an SVG of the mark; a reversed single-colour version; a horizontal lockup of the mark with the words "AI AgriBench"; a favicon; and a branded og:image. Until the lockup file exists, do not assemble one ad hoc.
The mark is a single ink and must not be redrawn, recoloured, rotated, outlined or placed inside another shape. Keep clear space of at least one leaf-width on all sides. Minimum size 32px on screen.
Contact
aiagribench@lists.illinois.edu
TODO (Committee): spokesperson policy. Who speaks for AI AgriBench in media interviews — the Steering Committee chair, a rotation, or by topic?