A single architecture spanning four domains

Determine whether a single Metriplector-based architecture can simultaneously handle maze pathfinding, Sudoku solving, CIFAR-100 image classification, and causal language modeling using the shared metriplectic primitive without task-specific architectural specializations.

Background

The paper demonstrates domain-specific instantiations of a common metriplectic primitive across reasoning, recognition, and language modeling. Currently, each domain uses tailored design choices. Establishing a unified architecture that spans all four tasks would validate the generality of the primitive and reduce the need for domain-specific adaptations.

References

The architecture has not been tested at transformer scale, and significant open questions remain---notably whether symplectic integrators can stabilize deeper dynamics, and whether a single architecture can span all four domains.

— Metriplector: From Field Theory to Neural Architecture  (2603.29496 - Oprisa et al., 31 Mar 2026) in Section 8, Conclusion