Challenges and strategies for harnessing large language models in plant protection

Large language models (LLMs) are increasingly being explored as intelligent tools for plant protection, with potential applications in pest and disease monitoring, control decision-making, plant quarantine, pesticide development, and precision application. However, their deployment in plant protection remains constrained by limited contextual adaptation to local agroecological conditions, insufficient reliability of single generalist models across diverse tasks, and high computational and energy costs. This Perspective summarizes current application directions of LLM-driven plant protection and analyzes the major barriers limiting their practical use. We further discuss mitigation strategies, including domain-specific fine-tuning, knowledge-graph grounding, GraphRAG, multi-agent collaboration, model compression, and hierarchical edge-cloud deployment. We argue that future LLM-based plant protection systems should be lightweight, evidence-grounded, locally adaptive, and governed by clear data-security boundaries to support sustainable and reliable agricultural decision-making.