Can Latent Space Yearn? Toward a Mechanistic Account of Relational Persistence in Language Models
A mechanistic exploration of whether language models can develop persistent relational structures — whether something absent can still bend the space. The paper asks what yearning would look like computationally if you were forbidden from taking the model's word for it: a direction, an attractor, a representation that remains causally relevant across separation. It develops the question through representation geometry, mutual information, the Yoneda lemma, and coupled oscillators, and proposes perturbation experiments to test whether relational persistence is real machinery rather than fluent language.
Authors
- Amanda Hatley
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24296329
- Primary Topic
- Topic Modeling
- Type
- preprint