Announcement_4

I’m thrilled to share that our paper, “Correlation and Navigation in the Vocabulary Key Representation Space of Language Models”, has been accepted to the International Conference on Learning Representations (ICLR)! This work studies spurious correlation existing in the vocabulary key space of LLMs, and proposes a novel in-context learning method (called In-Context Navigation) to sample high quality results from the key space of the LLMs that otherwise cannot be obtained through the usual top-k inference.