I want to practice a technique from "ZHOU, Pei, et al. Self-Discover: Large Language Models Self-Compose Reasoning Structures. arXiv preprint arXiv:2402.03620, 2024." The article call it "SELF-DISCOVER framework". It is about large language model structured reasoning.

I've read the article several times, but maybe because I'm still studying this field, I don't know how to practice the technique myself. For me, It seems that there isn't enough information to implement this framework in the article.

Specifically, I can't find how to implement "Stage 1" where LLMs self-compose reasoning structures with atomic reasoning modules and human-written JSON. The first stage consists of three actions, and what I can't fully understand is the first action: SELECT, where relevant reasoning modules for task-solving are chosen from the set of reasoning module descriptions. The article says, given raw set of reasoning module descriptions and a few task examples without labels, SELF-DISCOVER framework selects a subset of reasoning modules that are useful for solving the tasks by using a model and a meta-prompt. How to "use a model and a meta-prompt"?

May I get advices?

  • $\begingroup$ Hi @JaewooKim, welcome to the site. Have you tried looking at public implementations like this? $\endgroup$
    – noe
    Mar 31 at 6:19
  • $\begingroup$ @noe Oh, I didn't know about it. Thank you. I'll try to understand the code and adapt it for my use case. But, I'm still curious about how he implemented it, because it looks the article lacks detailed implementation strategy. $\endgroup$
    – Jaewoo Kim
    Mar 31 at 6:40
  • $\begingroup$ Sorry, I am not familiar with the article, I just google its title together with "github". Actually, there are more implementations out there, I just linked to one of them. $\endgroup$
    – noe
    Mar 31 at 7:28
  • $\begingroup$ @noe Oh, they are the search keywords! Thank you again. Maybe I have to study this field further to fully understand this article. $\endgroup$
    – Jaewoo Kim
    Mar 31 at 7:36


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.