Multimodal LLMs Hallucinate in Agricultural Image Tasks

A recent study published on arXiv (2605.27595) explores the phenomenon of hallucination in multimodal large language models (LLMs) within the context of agricultural imagery. The research focuses on two main aspects: the image-to-text process, where LLMs analyze crop or field images to identify conditions such as biotic and abiotic stresses, and the text-to-image process, where synthetic agricultural scenes are created based on prompts. The study identifies errors including biological inconsistencies, contextual inaccuracies, and agronomic implausibilities, assessed using domain-specific criteria across various imaging types. Both interpretive and generative tasks reveal recurring hallucination patterns, emphasizing the potential dangers of misleading agronomic insights.