Our paper on the geometry of language model embeddings received an Outstanding Paper Award at the Conference on Language Modeling (COLM).
The paper, led by Andrew Lee (Harvard, Wattenberg-Viegas Lab), characterizes the local and global geometry of language model token embeddings and finds similarities across language models. We introduce a simple measure of intrinsic dimension and demonstrate that embeddings lie on a lower dimensional manifold, and that tokens with lower intrinsic dimensions often have semantically coherent clusters, while those with higher intrinsic dimensions do not. We further introduce Emb2Emb, an approach for linearly transforming steering vectors from one language model to another, despite the two models having different dimensions.