meta-template pool retrieval and updating code

[GreenyWind]

请问这部分代码计划什么时候更新呢？很想在生物问题上试试。

[YangLing0818]

Thanks for your attention. We may update this part a month later.

[xiezhipeng-git]

最关键的部分代码没有看到。能介绍下到底是模型总结的思维模板，还是人类总结的？是训练或微调中总结的还是推理中总结的？和针对问题处理过的rag又有哪些不同呢？

[weizhehuang0827]

最关键的部分代码没有看到。能介绍下到底是模型总结的思维模板，还是人类总结的？是训练或微调中总结的还是推理中总结的？和针对问题处理过的rag又有哪些不同呢？

我也同样想知道这块的具体细节，好像作者文章的意思是在evaluation的时候同步更新meta buffer

[ZhaochenYu0201]

Thank you for your attention. Our current repository only contains the basic part of our BoT, which is the demo version. The core code, including the meta buffer library and buffer manager, has not been released yet. Please stay tuned for updates.

As for the thought templates, they are generated by LLMs. During the construction phase of the meta buffer, we prepare three general coarse-grained thought templates for utilization. These thought templates are further instantiated from the three coarse-grained templates based on tasks and are refined by the buffer manager to form new thought templates.

The RAG mainly extracts facts related to the corresponding problems from external sources to address hallucination issues, thereby enhancing the accuracy of LLMs in answering professional questions in specific domains. However, it cannot provide precise guidance for complex reasoning problems. Compared to the RAG processed for tasks, our thought templates contain general approaches to problem-solving and represent abstract high-level thoughts that could help solve difficult reasoning problems.

For more details, please refer to our paper.

[dbinthesky]

缺乏最关键的代码，现在的运行逻辑是调用了test_templates.py里预定义的解题模板。paper中的一个关键手段是Problem Distiller，从问题中自动化地提取template，这个在现在的代码中并没有看到。

[ZhaochenYu0201]

目前我们开源的是不包含meta-buffer library和buffer manager的版本，之后我们会对这两个模块逐一更新，请大家耐心等待！同时针对这个问题我们在README中也有说明，感谢大家的关注！

[Lbaiall]

请问这部分代码计划什么时候更新呢？很想在生物问题上试试。

@GreenyWind 你启发于哪种生物问题呢，我也是想在动物行为推理上去进行启发和猜想

[gaoxix]

Thanks for updating the core code about the buffer manager and meta buffer soon. I want to Leave two questions.
Firstly,now the thought templates is all about the program of thougts. So where is the difference between the PoT(progarm of thoughts) and the BoT.
Secondly, BoT has not feedback mechanism，so if through the reasoning process the answer is wrong，how to deal with the situation? or still save the high-level thoughts? how to address the more general ploblems?

[william4s]

目前我们开源的是不包含meta-buffer library和buffer manager的版本，之后我们会对这两个模块逐一更新，请大家耐心等待！同时针对这个问题我们在README中也有说明，感谢大家的关注！

@ZhaochenYu0201 It's been nine months, will there be any updates?

[ZhaochenYu0201]

@william4s Thank you for your continued interest in our work. We appreciate your patience, as we recognize it has been some time since our last update.

Our team is currently focused on addressing the challenges related to the buffer manager's generalization capabilities across diverse task domains. We are developing more robust methodologies to automatically extract and adapt thought templates for a wider range of applications, which will significantly enhance the system's versatility.

We plan to release updated code in the coming months as part of a new project that will addresses these generalization challenges. We apologize for the delay and encourage you to continue following this repository for announcements regarding these improvements.

Currently, we can offer some guidance for extracting templates for reasoning tasks. You may utilize the following prompt structure to develop a template library from previously solved problems:

"You are a reasoning expert and a teacher skilled in Feynman-style explanation. Please carefully analyze the Question and its corresponding Solution given below, provide a detailed thought analysis of the solution, extract key learning points, and then create a structured thought template..."

Thank you for your understanding and ongoing interest in our research.