Artificial Intelligence and Electrical & Electronics Engineering: AIEEE Open Access
Chiral Transformers: Directional Attention and Information Confirmation via Symmetry Breaking
Abstract
Chur Chin
We present a chiral modification of Transformer self-attention in which the parity symmetry of the standard attention operator is explicitly broken into right-handed and left-handed components. This decomposition induces a directional information flow that admits a natural interpretation as an advection–dissipation mechanism in a neural evolution equation. We establish well-posedness of the discrete chiral update, derive an energy-type estimate, and prove a stability theorem showing suppression of spurious amplification under mild operator norm bounds. The framework provides a mathematically controlled notion of information confirmation analogous to backward verification in convolutional networks and to dissipative corrections in partial differential equations.