Indirect expression is essential in many scenarios of real life. Enabling machines with the capability of indirect expression as humans is of great significance in many applications. Despite the advancements of large language models (LLMs), their capabilities to handle indirect expression remain limited. Against this problem, we propose a novel framework to improve LLMs’ performance through the lens of controllable text generation. Specifically, a dynamic direct word list is generated iteratively at first, which is critical for LLMs to generate the texts conforming to indirect expression. Subsequently, we design a constrained beam search algorithm to improve LLMs’ capabilities to avoid direct words. As a significant but barely studied task, rational metrics for assessing LLMs’ indirect expression capabilities are absent. Thus, we specially devise a credible evaluation protocol to evaluate LLMs’ performance on indirect expression. Furthermore, we construct a new dataset including Chinese and English samples, upon which our experiments have demonstrated our proposed framework’s effectiveness in improving LLMs’ indirect expression capabilities. The resources are available in the repository https://github.com/Haelles/Indirect .