The Router Within: Eliciting Native Skill Routing from a Frozen LLM

The Router Within. Eliciting Native Skill Routing from a Frozen LLM. By Ruishuo Chen and four others, from Tsinghua University. Published on the fourteenth of September. No code repository is linked to it.

Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise depends on choosing the right one. Deployed harnesses route by preloading every skill's metadata into the context. That disperses the agent's attention, and it caps how large the library can grow, since every new skill takes up room. Retrieval pipelines move the selection out of the context. But in doing so they also move it out of the agent's own capability, handing the choice to an external system.

The authors show that a frozen agent LLM already carries the routing signal in its own forward passes, and that two linear maps are enough to read it out, with no skill text in the context at all. Their method is Gavel, short for Glance And Verdict from a frozen LLM, and as the name says, it works in two steps.

The glance projects mid-layer states for the task and for each skill through the two maps, which are the only parameters trained. It scores the full library against compact per-skill banks, each built by a single forward pass when the skill is installed. That gives a shortlist cheaply. Because each bank is built once at installation, adding a skill costs one forward pass, and scoring it later needs no skill text in the context.

The verdict then resumes the forward passes of the shortlisted skills and reads the model's own likelihood and its own yes or no judgment. Those readings are fused with the glance as a product of experts, so the final choice has to satisfy both.

Trained once, Gavel transfers zero-shot to three public benchmarks and to SkillTraj, a new benchmark of three hundred and seventy-two simulated agent trajectories. On Qwen3-32B it outperforms progressive disclosure. It also outperforms retrieve-and-rerank pipelines that add between one point two billion and sixteen billion external parameters. The margin reaches thirteen point four points on written tasks, and twenty-one point nine when the need for a skill arises in the middle of a rollout.

Routing accuracy improves as the backbone improves. And in a bash-agent harness, the same thirty-two billion parameter model triggers the correct skill on Skill-Use more often than far larger frontier models running in Codex.

Two contrasts frame the result. Against preloading metadata, Gavel keeps the context clear. Against retrieval, it keeps the decision inside the agent's own forward passes, rather than delegating it to billions of external parameters. The claim underneath is that the model already knows which skill it needs, before any skill text is shown to it. Reading that knowledge out with two small maps keeps routing inside the agent, adds almost nothing in parameters, and lets the library grow without crowding the context.