AgriChain: Visually-Grounded Expert-Verified Reasoning for Interpretable Agricultural Vision–Language Models
Summary
Vision–language models (VLMs) can often name a plant disease but rarely explain the diagnosis. They are also weakly grounded in agronomic imagery and tend to be overconfident. Existing agricultural resources such as PlantVillage, AgroInstruct and AgroMind provide labels or QA pairs but no step-by-step rationales.
The authors built AgriChain, about 11,000 leaf images merged from PlantVillage, PlantDoc and PlantCLEF. The images were de-duplicated and mapped to 33 disease classes plus a healthy class. For each image, GPT-4o drafted a symptom-based explanation. A professional agricultural engineer then corrected it, standardized the descriptors (lesion color, margin, distribution) and assigned a High/Medium/Low confidence level.
Fine-tuning Qwen-2.5-VL-3B on these image–rationale pairs produced AgriChain-VL3B, which outputs both a diagnosis and an expert-style justification. On a 1,000-image test split it reached 73.1% top-1 accuracy, roughly 17 points above the best commercial baseline. Its explanations consistently cite the relevant visual cues.
The work shows that a compact model supervised with verified reasoning can become a more accurate and auditable diagnostic aid. The authors frame this as relevant to food security in water-scarce regions such as MENA.
Key points
- Data splits: AgriChain has ~11k images across fruits, vegetables and other crops (apple, grape, citrus, tomato, potato, pepper, banana, strawberry). It is split 9,000 / 1,000 / 1,000 for train, validation and test, with every class present in each split.
- Annotations: each image carries a disease label, a calibrated confidence score and an expert-verified chain-of-thought rationale.
- Results: AgriChain-VL3B achieves 73.1% accuracy and weighted-F1 = 0.655. The best baseline scored 55.8% (Gemini Pro); the other baselines were Gemini 1.5 Flash and GPT-4o-Mini.
- Inconsistencies in the paper: macro-F1 is reported as 0.466 in the abstract but 0.676 in the conclusion. The drafting model is named both GPT-4o and GPT-4o-mini.
- Released resources: the authors also propose AgriReason-Bench with a Region–Text Alignment (RTA) faithfulness metric. Dataset and code are on GitHub (hazzanabeel12-netizen/agrichain).
Cite
@inproceedings{mahmood2026agrichain,
title = {{AgriChain: Visually-Grounded Expert-Verified Reasoning for Interpretable Agricultural Vision–Language Models}},
author = {Mahmood, Hazza and Yu, Yongqiang and Anwer, Rao},
booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
year = {2026},
doi = {10.63317/2bv5k9hnduop},
url = {https://doi.org/10.63317/2bv5k9hnduop}
}References
- Jean-Baptiste Alayrac et al. 2022. Flamingo: A visual language model for few-shot learning. find our card
- Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023a. Qwen-vl: A versatile vision-language model for understanding and generation. find
- Y. Bai and et al. 2023. Evaluating domain adaptation for multimodal models. find
- Yue Bai et al. 2023b. Qwen-vl: A versatile vision– language model for understanding and generation. find
- Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natural language inference with natural language explanations. find our card
- Google DeepMind and Google AI. 2024a. Gemini 1.5 flash: High-speed multimodal reasoning. find
- Google DeepMind and Google AI. 2024b. Gemini 1.5 pro: Scaling multimodal understanding and reasoning. find
- Hervé Goëau, Alexis Joly, Pierre Bonnet, and et al. 2022. Overview of plantclef 2022: Plant identification and disease recognition challenges. find
- Thomas Lanham, Samuel R. Bowman, and Ethan Perez. 2023. Measuring faithfulness in chain-of-thought reasoning. find our card
- Junnan Li et al. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. find our card
- Y. Liu, D. Iter, Y. Xu, and et al. 2023. G-eval: Nlg evaluation using gpt-4 with better human alignment. find
- P. Lu, S. Mishra, T. Xia, and et al. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems (NeurIPS 2022). find
- OpenAI. 2024. Gpt-4o mini: Efficient multimodal large language model. find
- Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamila Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. find our card
- Roboflow Universe. 2022. Leaf disease dataset. find
- Dhruv Singh et al. 2020. Plantdoc: A dataset for visual plant disease detection. find
- Jason Wei and et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. find our card
- Bo Yang, Yunkui Chen, Lanfei Feng, Yu Zhang, Xiao Xu, Jianyu Zhang, Nueraili Aierken, Runhe Huang, Hongjian Lin, Yibin Ying, and Shijian Li. 2025. Agrigpt-vl: Agricultural vision–language understanding suite. find
- Yuchen Zhang, Ming Li, Rui Zhao, et al. 2024. Multimodal chain-of-thought reasoning in vision–language models. find
- Chenyang Zhao et al. 2025. Planning with chain-of-thought in embodied and multimodal agents. find
- Hughes and Salathé. 2015. PlantVillage Dataset. PlantVillage Project, 1.0. High-quality leaf images collected under controlled environments. find
This page in Markdown → https://md.research.elamur.ai/free/papers/mahmood-2026-agrichain-visually-grounded-expert-verified/