Spectral Routing: Context-Dependent Expert Selection via Optical Refraction in Neural Networks

We present Spectral Routing, a context-dependent routing mechanism for Mixture-of-Experts (MoE) language models inspired by optical refraction. Each token query carries a spectral color vector encoding conversational context. Semantic nodes act as prisms with learned refractive indices, and Snell's law determines routing angles. This enables the same node to produce different routing decisions for different contexts, resolving token polysemy without duplicating expert weights. Our full pipeline — combining Snell refraction, chromatic aberration (multi-band decomposition), total internal reflection (domain boundaries), and phase-coherent interference — achieves 98.4% polysemy resolution (435/442 correct, 80 polysemous words across 442 context pairs), outperforming the standard linear MoE gate (72.3%) by 26 percentage points. Computational overhead is less than 0.12% of base BVH traversal cost. Total internal reflection provides emergent domain rejection without learned parameters, and Snell's law offers geometric regularization (energy conservation, reversibility) that a generic context-conditioned MLP lacks. Validated on OLMoE-1B-7B using an NVIDIA RTX 5070 Ti. This work extends the SpectralAI routing system (DOI: 10.5281/zenodo.19457288).