Large Language Models (LLMs) have demonstrated remarkable zero-shot capabilities across various domains. This paper pioneers the application of LLMs’ outstanding knowledge and reasoning abilities to the challenging task of Traffic Matrix (TM) imputation. However, the application poses significant challenges due to the skewed TM distribution and the deficient traffic feature under low sampling rate. To address these issues, we propose TM-LLM, the first LLM-based model specifically designed for TM imputation. Our approach includes two critical designs: Firstly, we develop an adversarial training strategy to pre-impute TM data, allowing the LLM to understand the distributional features even when faced with extensive missing data. Secondly, we devise a TM-specific embedding scheme along with a crafted prompt template, which enables our approach to harness LLMs’ exceptional inferential ability. Experimental results show that TMLLM significantly outperforms state-of-the-art imputation methods, achieves a notable 16.5% -44.8 % improvement in accuracy over the current best baseline, while reduces measurement costs by 80 % - 96 %. It can accurately capture the traffic pattern even when the sampling rate is extremely low. The code for reproducing our experiments is publicly available1. These findings strongly indicate the breakthrough potential of LLMs in network TM analysis tasks.1The experimental codes with our methods and the datasets are available at https://github.com/FILingK/TM-LLM