## Introduction Large language models have become increasingly popular for their ability to understand and generate complex texts. However, this increased complexity has also a downside: the length of the generated text can become too long, reducing the efficiency of the model and making it less accessible. To solve this problem, researchers have proposed a new technology called Leash. This framework uses an automatic learning approach to dynamically control the length of the generated text. ## How Leash works Leash is based on a concept called "penalty" that represents the penalty applied to the model when its responses are too long. This penalty is calculated dynamically based on the task requirements and can vary in intensity depending on the length of the response. In this way, Leash helps the model to produce more concise and effective responses without sacrificing its reasoning ability. ## Experiments and results Researchers have tested Leash on two very large language models: Deepseek-R1-Distill-Qwen-1.5B and Qwen3-4B-Thinking-2507. The results were impressive: Leash reduced the average length of the generated text by 60% without sacrificing the model's reasoning ability. ## Conclusion The technology Leash represents a significant turn for large language models. With its ability to dynamically control the length of the generated text, Leash helps improve efficiency and concision, making the model more accessible and useful for a wider audience. ## Resources To learn more about the Leash technology, you can visit the researchers' website or read scientific articles on the topic.