Performance Analysis and Enhancement of Multimodal Large Language Model Phi-3.5 for Plant Disease Recognition

Large Language Models (LLMs) show significant potential across many professional fields. In the agricultural domain, multimodal LLMs can provide farmers with effective support for recognizing and diagnosing plant diseases. Assessing LLM performance in plant disease image analysis is of practical importance for improving plant disease recognition systems. This study evaluates the classification performance of Phi-3.5, a multimodal LLM released by Microsoft, on the PlantVillage dataset. Experimental results indicate that although LLMs often exhibit “hallucinations,” erroneous or imprecise outputs, when analyzing plant disease images, careful prompt engineering and LoRA fine-tuning can significantly enhance their performance. Our findings highlight the potential of LLMs in plant disease recognition. With ongoing refinements, these models are expected to play a critical role in practical applications.