Abstract
Polycentric Embedding Fields (PEF) introduce a non-monolithic representation paradigm for semantic encoding in which each input instance is modeled not as a single vector in a latent space, but as a structured field composed of multiple interacting semantic centroids, each capturing a distinct interpretive axis of meaning. This formulation directly addresses limitations inherent in conventional embeddings, including semantic averaging, loss of intent granularity, and failure under polysemy or compositional ambiguity. The disclosed architecture encodes documents and queries as distributions over latent subspaces with explicit uncertainty modeling, enabling selective alignment between semantically compatible facets rather than forcing global similarity. By transforming embedding from a point estimate into a parameterized field with internal structure, PEF provides a mathematically grounded and computationally realizable mechanism for improving retrieval precision, reducing hallucination propagation in retrieval-augmented generation systems, and enabling fine-grained semantic governance at scale.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Manuel-Devadoss, Johny, "Polycentric Embedding Fields (PEF)", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9961