Abstract
We propose Hierarchical Resonance Field Networks (HRFN), a sequence modeling framework that replaces discrete token-level attention with a continuous-time dynamical system defined over a hierarchical latent field. Instead of computing pairwise interactions via similarity in embedding space, HRFN defines interactions as phase- and frequency-coupled resonance dynamics over a structured manifold.
Formally, each input token is mapped to a localized field excitation, and computation proceeds via iterative evolution of a sparse interaction graph induced by spectral alignment constraints. The resulting architecture generalizes self-attention, recurrence, and memory retrieval as special cases of a unified resonance operator acting over latent fields. We show that under bounded spectral sparsity assumptions, the model reduces computational complexity from quadratic to near-linear while preserving global dependency modeling via multi-hop resonance propagation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Manuel-Devadoss, Johny and Prasad, Deepthi, "Hierarchical Resonance Field Networks (HRFN)", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9996