You're right, the preprint was submitted ~1.5yr ago. But the paper was revised on 16 Apr 2023. I picked this one up in November 2023 as I came across the revision.
I add prompt examples on the papers I simplify but in this one I couldn't yet -- its still on my backlog to either add more info to this article or to archive it.
But I want to bring the focus here to the method rather than the LLM it was tested on. If the method can be better than CoT for GPT-3, it can serve well for the open-source LLMs that are currently at par or better than GPT-3 (like LLAMA, Mistral, etc.).
https://arxiv.org/abs/2205.10625
This is the only paper I could find about this. Are you going through relatively older papers and summarizing them, or did I miss a more recent paper?
Goodness, it's so crazy that this paper seems pretty old to me because its preprint was submitted ~1.5yr ago and it doesn't use GPT-4
You're right, the preprint was submitted ~1.5yr ago. But the paper was revised on 16 Apr 2023. I picked this one up in November 2023 as I came across the revision.
I add prompt examples on the papers I simplify but in this one I couldn't yet -- its still on my backlog to either add more info to this article or to archive it.
But I want to bring the focus here to the method rather than the LLM it was tested on. If the method can be better than CoT for GPT-3, it can serve well for the open-source LLMs that are currently at par or better than GPT-3 (like LLAMA, Mistral, etc.).
Jz
I I'm