Ideas

This is a updated version of my previous post on stochastic language models. This post can be read alone, since I have modified the previous idea significantly. Simply put, I still believe a stochastic language model is more desirable than a constant language model, but my previous immature post is subject to a few fallacies.

Considering language models as random variables brings insights into the challenges in LLMs and the mechanisms of various augmentation methods to improve LLMs on hallucination, biases, continual fine-tuning/knowledge editing/instruction fine-tuning/RLHF, tool learning, role playing, retrieval augmentation…