[D] Replicating the inner layers of LLMs with weight-sharing

Disclaimer: I don't have the means to train larger models, but have a lot of curiosity. Let me apologize for discussing cheap ideas without putting in the hard work to try these ideas out.

LLMs have issues handling multi-step or recursive reasoning. Their only way to solve similar problems is to *talk their reasoning through*.

For instance they know what "the continent south of Europe" is, as well as what "the southernmost country in Africa" is. But they struggle at …

africa country curiosity europe ideas llms machinelearning reasoning recursive talk work