Text Has Curvature
A new paper claims language itself is curved, not just our models of it.
Deep Dive
A new arXiv paper titled 'Text Has Curvature' argues language has an intrinsic, measurable curvature native to text itself, not just an artifact of embedding spaces. The researchers propose 'Texture,' a word-level discrete curvature signal defined via a Schrödinger bridge. They demonstrate its utility by improving long-context inference through curvature-guided compression and boosting retrieval-augmented generation via curvature-guided routing, establishing a new text-native curvature paradigm.
Why It Matters
This could lead to more efficient and accurate language models by using curvature as a built-in guide for tasks like compression and information retrieval.