The field of large language models (LLMs) is moving towards developing more efficient reasoning methods. Researchers are exploring ways to reduce the computational overhead associated with lengthy reasoning content, such as chain-of-thought (CoT) reasoning, without sacrificing accuracy. Various approaches are being proposed, including early exit methods, pruning techniques, and compression algorithms. These innovative methods aim to improve the efficiency of LLMs while maintaining their performance. Noteworthy papers in this area include FlashThink, which introduces a verification model to identify the exact moment when the model can stop reasoning, and R1-Compress, a two-stage chunk-level compression framework that preserves local information and coherence. Additionally, papers like DRP and ThinkSilently highlight the importance of aligning reasoning structures with model capacity and exploring diverse reasoning paths to achieve efficient and accurate reasoning.