With the proliferation of multimodal information on social media, the task of multimodal aspect-based sentiment classification (MA-BSC) gained significant attention. However, existing methods for this task heavily relies on the large-scale supervised data, which is costly and time-consuming to obtain. Recent efforts have shifted towards few-shot MABSC, where small-scale pretained models (e.g. BERT, BART) are fine-tuned using a limited samples ( \(1\%\) of the training set). While large language models (LLMs) have shown strong zero-shot and few-shot capabilities across various NLP tasks, their exploration in few-shot MABSC remains insufficient. To address this gap, we propose a novel fine-tuning framework for LLMs in few-shot MABSC, incorporating three key modules: few-shot sampling, label-introduced Chain-of-Though (CoT) Generation and CoT enhanced fine-tuning. Specifically, we first construct a few-shot training set via selecting informative samples based on the LLM’s zero-shot capabilities. Moreover, we generate CoT reasoning through the LLM’s zero-shot reasoning to guide fine-tuning, enhancing both performance and interpretability. Extensive experiments on two benchmark datasets (Twitter-2015 and Twitter-2017) demonstrate the superiority of our approach.