Vortex-Gated Mixture of Experts: Symbolic Zero-Parameter Routing via TERA Projection
Year: 2026
Type: article
Abstract: Mixture-of-Experts language models route each token through a small subset of 'experts' via a learned softmax gate, producing routing decisions that are opaque and difficult to audit. We propose Vortex-keyed routing: a closed-form, deterministic, differentiable gating mechanism in which the hidden state is projected into a four-dimensional semantically-labeled vector (the TERA register: Temporal, Emotional, Rational, Archetypal) and resolved to one of 16 named experts via a fixed binary mask (the Meji bit-mask). The resulting router is interpretable by construction, deterministic, and produces an auditable lineage of routing decisions per token. The mechanism is differentiable via sigmoid relaxation with a straight-through estimator. We describe the architecture, training procedure, soft-routing variant, properties relevant to alignment auditing, and outline a benchmark suite for empirical validation. Experimental results are preliminary; we are running the proposed evaluations and will report in v1.1.
Author: Whitehead, Weslyn Cory
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7160979895