The application of large language models (LLMs) was studied in evaluating the quality of Chinese metaphorical sentences. Building upon prior research and insights from cognitive linguistics, a multi-dimensional evaluation framework tailored to Chinese metaphors was developed. A high-quality human-annotated dataset was constructed based on the framework to serve as a benchmark for validating LLM performance in metaphor assessment tasks. Guided by conceptual metaphor theory, an LLM-based evaluation pipeline was proposed that integrates multi-turn dialogue and chain-of-thought prompting. Experiments were conducted to test the model’s effectiveness in two distinct tasks: direct scoring of metaphor quality and pairwise comparison for selecting superior metaphors within groups. Results demonstrate strong alignment between LLM evaluations and human judgments. In direct scoring tasks, the Pearson correlation coefficient reached 0.807, and for within-group selection tasks, Cohen’s Kappa coefficient of 0.831 was achieved. The proposed evaluation pipeline integrated conceptual metaphor theory with LLMs and achieved strong results on Chinese metaphor assessment.