Search

New AI test reveals plant disease chatbots aren't ready yet

Sakib SN, Haque N, Hossain MZ, Arman SE.

Crop Improvement

The AI plant-identification app on your phone might confidently misdiagnose the black spots on your tomato leaves, and this research shows even the most advanced vision AI models fail at reasoning through plant disease questions the way a real expert would.

Scientists gathered photos and questions from 45 different plant disease datasets to create a huge test, over 765,000 questions and answers, covering 38 crops and 89 diseases. They used it to check whether AI systems that can 'see' photos and answer questions actually understand plant diseases or just memorize patterns, and found that even top AI models struggle badly. The good news is that a smaller, more efficient AI model got much better after training on just a small slice of this new dataset, suggesting there's a real path toward AI garden helpers that can accurately diagnose sick plants.

Key Findings

1

Dataset contains 765,186 question-answer pairs across 150,841 images spanning 38 crop species and 89 disease conditions, compiled from 45 open-source sources including PlantVillage

2

Current frontier vision-language models, including recent open-source instruction-tuned multimodal LLMs, perform poorly on the benchmark

3

Fine-tuning a compact 2B-parameter model on a small fraction of the dataset produced substantial improvements across all 9 question categories

chevron_right Technical Summary

Researchers built a massive dataset of 765,186 plant disease Q&A pairs to test whether AI vision-language models can actually reason about crop diseases from photos, not just classify them. Current top AI models perform poorly on this test, but a smaller model fine-tuned on just a fraction of the data showed major improvement.

description

Abstract Preview

Original paper

A Visual Question Answering Dataset for Benchmarking Vision-Language Models in Plant Science.

Existing plant-disease datasets target classification and detection, leaving vision-language models unable to support interactive, reasoning-based diagnosis. To address this, we present PlantExpert...

open_in_new Read full abstract

Abstract copyright held by the original publisher.

hub This connects to 8 other discoveries — crop-improvement, plant-signaling, urban-ecology 5 related articles

Was this useful?

mail Weekly plant science — one email, Saturdays.

Share: X/Twitter Reddit
arrow_forward Next Discovery

Gene editing removes 97% of celiac-triggering proteins from bread wheat

It could mean that people with celiac disease — roughly 1 in 100 worldwide — may one day safely eat bread made from real wheat, without sacrificing the taste...

sensors Plant Signaling
Topic
sensors

Plant signaling encompasses the molecular and cellular mechanisms by which plants perceive and respond to environmental changes, hormonal signals, and stress conditions. These signaling pathways regulate fundamental biological processes including growth, development, nutrient acquisition, and

arrow_forward Explore topic