Question about state of research

[isamu-isozaki]

Hi! Thanks for the repository. I made a blog going over some of these papers here and my general conclusions were

1. LLMs lack understanding of complicated relationships between characters so they can't make say a mystery
2. LLMs have a forgetting-in-the-middle problem
3. For pacing/suspense etc, the recursive prompting strategy kind of works but it's more expensive and you have to develop a prompting strategy for each error in the LLMs capability. This can be optimized by training using like SFT+DPO
4. There are some foundational model training of LLMs for creative writing
   Do you think that's a correct conclusion atm?

[yingpengma]

Hey there,

Spot on with your conclusions—appreciate the contribution! I'll add your blog link to our resources.

Best,
Yingpeng

[isamu-isozaki]

@yingpengma tnx!