# Introduction A new approach to integrate research and reasoning in large language models. # Problem Language models are often optimized to improve their performance through semantic similarity-based knowledge retrieval or reasoning enhancement. This approach can be limited and not always effective in providing informative and creative responses. # Solution A new method introduces a strategy of knowledge retrieval that focuses on the logical structure of conversations. An MCTS algorithm is used to navigate through knowledge and find relevant information for the conversation context. # Coarse-to-Fine Approach The method follows a coarse-to-fine approach, consisting of two main phases: 1. **Sub-Region Identification**: The approach identifies a relevant sub-region of knowledge for the conversation context. 2. **Refined Search**: The search is refined within the sub-region to extract specific and pertinent information for the reasoning process. # Experiments The approach has been tested on two multi-turn dialogue datasets, with promising results. Models using this approach show improved alignment with human reasoning and increased diversity in responses. # Conclusion The new method introduces an innovative approach to integrate research and reasoning in large language models. This approach promises to significantly improve model performance and provide more informative and creative responses.