SCX Router is a lightweight GLiClass‑based router that assigns a suitability score to each inference‑time model label without performing autoregressive generation. It uses a 0.6B‑parameter checkpoint that couples a Qwen3 decoder with a shallow bidirectional scorer, forming a decoder‑KV execution path: a text‑only key‑value cache is kept across a session, only new dialogue turns are encoded, and transient candidate‑label tokens are evaluated without being written to the persistent cache. The same checkpoint also predicts task type, difficulty, reasoning mode, and expected output length, and it supports custom zero‑shot labels.
For task generation we built a task ontology comprising 23 families, 115 task types, 345 routable subtypes, 1,173 synthetic examples, and an orthogonal axis of 30 domains. Using this structure we generated 150,000 verifier‑scored tasks and 15,000 open‑ended tasks, then trained the Qwen3 decoder on them while explicitly separating learned request prediction from per‑task policies such as eligibility, cost, cache reuse, safety, and sovereignty.
Across six LiveBench subsets the router outperforms the mean candidate; on a selected 1,000‑task subset it achieves an aggregate top‑1 score of 0.707 versus 0.696 for the strongest fixed model, with benchmark‑dependent gains.
Review