A new benchmark makes agricultural AI prove it can both name a crop disease and point to exactly where it is, not just name it.
Researchers built AgroGround by converting labels from eight existing agricultural visual question answering datasets into 794,850 instruction examples that tie each disease or object label to the specific image region it describes. Healthy images were deliberately included as negative examples, teaching models to return nothing when asked about disease on a plant that has none. The team fine-tuned a vision-language model on a mix of plain localization instructions and combined recognition-and-localization instructions, then tested it on 1,480 human-verified images held out from training. Training on localization alone actually hurt accuracy, dropping it from 51.8% to 29.1%, but adding the combined instructions pushed recognition to 72.6% and raised joint accuracy, naming and locating the problem correctly together, from 19.2% to 43.3%.
Most agricultural AI demos show a model confidently naming a disease, but a name without a location is not much use to someone deciding where to spray or cut. This result is a useful reality check: a model that is good at recognition is not automatically good at localization, and training for both at once is what actually closes the gap. The healthy-image test matters just as much, since a system that correctly says nothing is wrong 95.0% of the time is more useful in a field than one that always finds something to flag.
The resulting model, built on a 2B parameter base, reportedly outscored the very pipeline that generated its training labels on grounding accuracy, a reminder that in this field the automated label makers are still far from perfect themselves.