The Generation Half: Why Prompt-Routing Studies Understate Domain Specialization in Mixture-of-Experts Models. Preprint, version 1. 0 (June 2026). Not peer reviewed. A widely cited reading of mixture-of-experts (MoE) language models holds that their experts do not specialize by subject matter: routing measured over supplied text spreads each expert across many topics. This paper shows that the reading rests on half of the forward pass. A transformer routes tokens in two regimes, a parallel prefill pass that reads the prompt and a step-by-step generation pass that writes the answer, and supplied-text routing is the prefill half. Across a suite of prompts spanning twenty academic domains, we identify the expert that receives the most routing weight in each domain, separately for each pass. Over prefill a single high-frequency expert wins almost every domain (18 of 20 in a 35B model, 13 of 20 in a 122B model) and routing looks domain-blind. Over generation the winners disperse into distinct per-domain experts, and the winner-distribution entropy moves from 0. 13 to 1. 00 (35B) and from 0. 42 to 0. 95 (122B). The shift is large and one-directional, survives a control that pads all prompts to a common length, and reproduces in the second, larger model whose expert indices are entirely different while the prefill-concentrated, generation-dispersed pattern holds. Domain specialization in these models is largely a property of the generation pass, and routing read off prefill, or pooled across passes, understates it. This deposit contains the paper (LaTeX source and built PDF), the three figures and the script (makefigures. py) that reproduces them from the winner lists without re-running any model, a verified bibliography, a claim-by-claim source index (SOURCES. md), and the supporting data/ directory: three runs (a 35B 60-prompt primary probe, a 35B token-balanced 3-chunk control, and a 122B replication) with per-domain winner tables and raw router tensors, each mapped to the claims it supports in data/README. md.
Jeffrey W. Shorthill (Sun,) studied this question.