Toward Manifest Relationality in Transformers via Symmetry Reduction
arXiv:2602.18948 · doi:10.1103/v57m-nhhq
Abstract
Transformer models contain substantial internal redundancy arising from coordinate-dependent representations and continuous symmetries, in model space and in head space, respectively. While recent approaches address this by explicitly breaking symmetry, we propose a complementary framework based on symmetry reduction. We reformulate representations, attention mechanisms, and optimization dynamics in terms of invariant relational quantities, eliminating redundant degrees of freedom by construction. This perspective yields architectures that operate directly on relational structures, providing a principled geometric framework for reducing parameter redundancy and analyzing optimization.
V2, 14 pages, 2 columns, Accepted for publication in APS Open Science
References in corpus (11)
- Geometric relational framework for general-relativistic gauge field theories
- Covariant canonical formulations of classical field theories
- Dressing vs. Fixing: On How to Extract and Interpret Gauge-Invariant Content
- On the Meaning of Local Symmetries: Epistemic-Ontological Dialectics
- Dressing fields for supersymmetry: The cases of the Rarita-Schwinger and gravitino fields
- The dressing field method for diffeomorphisms: a relational framework
- Off-shell supersymmetry via manifest invariance
- Reassessing the foundations of Metric-Affine Gravity
- Unconventional Supersymmetry via the Dressing Field Method
- Raising galaxy rotation curves via dressing
- Spacetime boundaries do not break diffeomorphism and gauge symmetries