Train a vision model on agricultural images and it learns to associate leaf spots with diseases. But which diseases? The training data — scraped from the web, labeled by non-experts — contains biological hallucinations. A brown spot gets labeled as bacterial blight when it's actually sunscald. The model learns the wrong association with high confidence.
AgriChat (arXiv:2603.16934) introduces the Vision-to-Verified-Knowledge pipeline: visual captioning generates a description, then web-augmented scientific retrieval grounds that description in peer-reviewed phytopathological literature. The model doesn't just see the image — it checks the image's interpretation against published science before accepting it as training data.
The result is the AgriMM benchmark: over 3,000 agricultural classes, 607,000 visual Q&As spanning species identification, disease recognition, crop counting, and ripeness assessment. AgriChat scores 77.4% on its own benchmark and transfers to unseen datasets (PlantVillage: 74.3%, CDDM: 69.9%) without fine-tuning.
The mechanism that matters: the pipeline doesn't fix the model. It fixes the training data. Standard fine-tuning assumes the corpus is accurate — a dangerous assumption in any domain where visual similarity doesn't map cleanly to categorical identity. A healthy leaf and a nutrient-deficient leaf can look identical to a camera. The difference lives in the literature, not the pixels. Grounding in text isn't a supplement to vision. It's a correction of vision's endemic overconfidence.