The Router Within — Glance-and-verdict skill routing — What does it mean?

The news. On 14 September 2026, researchers posted "The Router Within: Eliciting Native Skill Routing from a Frozen LLM", describing a method they call Gavel. The claim is that the agent model you are already running knows which skill the task needs, and that two learned linear maps suffice to read that judgement out of its hidden states. Trained once, Gavel transfers with no retraining to three public benchmarks and to SkillTraj, the paper's own set of 372 simulated agent trajectories. On a Qwen3-32B backbone it reportedly beats progressive-disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters — by up to 13.4 points on written tasks and up to 21.9 points when the need for a skill only surfaces mid-rollout. Read the paper →

Stay with the concierge for a moment. The binder on the desk is what most agent harnesses do today: before the agent can choose anything, every skill's description is copied into its context window. That is fine while the library is small. As it grows, the paper says, it does two bad things at once: it spends context the task itself needed, and it disperses the agent's attention across descriptions that are mostly irrelevant to the task at hand. Gavel's starting claim is that the concierge recognised the right card the moment the request was spoken, and the harness simply never asked.

What "recognised" means here is concrete. As the frozen model reads the task, it builds a mid-layer state — the partly-processed vector sitting midway up its stack of layers. Every installed skill has one too, computed once by a single forward pass at installation time and kept as a compact skill bank. The glance projects both through the two learned linear maps into a shared space and scores the entire library against those stored banks. It compares vectors, not text, which is why nothing about the skills has to enter the context at all.

The glance only shortlists. The verdict then resumes the shortlisted skills' forward passes and reads two signals the model itself produces: its own likelihood for that skill, and its yes-or-no judgement about it. Those are fused with the glance score as a product of experts — the scores multiply, so a skill has to look right on both readings to win. The only parameters trained in the entire method are those two linear maps; the agent model itself is never touched, which is the work the word "frozen" is doing in the title.

That is also the difference from the outside booking agency. A retrieve-and-rerank pipeline keeps skill text out of the context too, but it hands the decision to a separate model. The paper is blunt about the trade: retrieval moves selection out of the context, "but also out of the agent's capability." Gavel's routing accuracy instead improves as the agent backbone improves — which the paper reports directly, and which follows from the router being the backbone.

Put a number on the outside-agency cost. Hold two things fixed: the agent is Qwen3-32B, and the skill library is whatever size it is — library size does not enter this arithmetic. The baselines Gavel is measured against bolt on a separate router; the paper puts those at 1.2B to 16B external parameters. At the small end that is 1.2 / 32, or about 3.8% more parameters to host. At the large end it is 16 / 32 — a second model half the size of the agent itself, loaded, served and carried on the cost profile purely to decide which skill to open. Gavel's side of the ledger has no third model on it: two linear maps, plus one forward pass per skill at installation — a one-off cost that does not repeat per request. What that buys, on the paper's reported numbers, is up to 13.4 points on written tasks and up to 21.9 points mid-rollout against exactly those pipelines.

Goes deeper in: AI Agents → Tool Use → Skills as a Primitive