The role of large language models in agriculture: harvesting the future with LLM intelligence

Significant accomplishments in many agricultural applications during the past decade attest to the fast progress and use of deep learning and machine learning methods in agricultural systems. However, these conventional models have a few drawbacks: They are not generalizable since they are trained on large, costly labeled datasets, require expert expertise to create and maintain, and are often built for specific applications. Significant accomplishments in language, vision, and decision-making tasks across several domains have been shown recently by massive pre-trained models, also known as large models (LMs). Recent years have seen large language models (LLMs) demonstrate remarkable competence in a variety of fields, including natural language processing (NLP), by encompassing different advancements in terms of architecture, training methods, context duration, fine-tuning, multi-modality, datasets, efficiency, benchmarking, and many other. The massive amounts of data used to train these models span many domains and modalities. After training, they can handle a wide range of tasks with less tweaking and less task-specific labeled data. Despite its effectiveness and promising future, agricultural artificial intelligence (AAI) has received less attention than other applications of LLMs. To better understand the problem area and open up new research pathways in this sector, this work aims to examine the possibilities of LLMs in smart agriculture by offering conceptual tools and a technical base. Herein, we delve into the potential applications of large models in agriculture, primarily categorizing them into four categories: Agricultural applications of large language models (LLMs), large vision models (LVMs) for precise agricultural applications, multimodal large language models (MLLMs) and model assessment, and intelligent and precise agriculture using reinforcement learning large models (RLLMs). Further, we review some of the most prominent LLMs, including three famous LLM families (GPT, LLaMA, PaLM), and discuss their characteristics, contributions, and limitations. Next, we evaluate famous LLM evaluation metrics and look at datasets for training, fine-tuning, and evaluation. Finally, we focus our discussion on issues and possible future research directions of LLMs in the agricultural sector. This review article aims to provide academics and practitioners with a panoramic perspective of the field and a quick reference to help them draw out relevant ideas from the extensive summaries of prior publications to broaden their LLM research.