The recent advancements in the field of large language models (LLMs) have been focused on enhancing their reasoning capabilities, particularly in complex and multi-step tasks. Researchers are exploring various strategies to improve the logical reasoning and adherence to complex instructions of LLMs, moving beyond the traditional chain-of-thought prompting. One significant trend is the use of knowledge distillation and intermediate-sized models to transfer reasoning abilities from larger models to smaller ones, addressing issues of data quality and soft label provision. Another notable direction is the development of frameworks that enable LLMs to self-correct and refine their responses to better meet specified constraints, leveraging tools for verification and refinement repositories for diverse constraints. Additionally, there is a growing interest in preference-guided reasoning and recursive learning approaches that allow models to iteratively improve their reasoning through self-teaching and feedback loops. These methods not only enhance the accuracy and efficiency of reasoning but also demonstrate the potential for smaller models to achieve similar performance levels as larger ones through innovative training techniques. The field is also witnessing advancements in aligning LLMs with multi-branch and multi-step preference trees, which offer more comprehensive preference learning and fine-grained optimization. Overall, the current research landscape is characterized by a shift towards more sophisticated and adaptive reasoning strategies that aim to make LLMs more reliable and versatile in handling complex tasks.