Morning Edition · Friday, July 24, 2026Published at 1:33 AM EDT · New York

Two Papers Probe the Inside of Mixture-of-Experts Routing

One frames expert selection as a frequency-driven code over chains of thought, the other uses expert-aware contrastive decoding to reduce hallucination.

Two arXiv preprints examine the routing logic inside Mixture-of-Experts (MoE) models, the architecture behind most efficient frontier systems, where a gate sends each token to a small subset of experts. The first, asking whether MoE routing…