Skill libraries let an LLM agent act beyond its parametric knowledge, but only if the agent picks the right skill. Common harnesses preload every skill's metadata into the context, which spreads the model's attention thin and limits how large the library can grow. Retrieval pipelines move selection outside the context, but also outside the agent's own capability.
A new paper, The Router Within: Eliciting Native Skill Routing from a Frozen LLM, argues that the frozen agent model already carries the routing signal in its forward passes. The proposed method, Gavel (Glance And Verdict from a frozen LLM), reads that signal with two linear maps and no skill text in the context.
Gavel works in two steps. A glance projects the task's and each skill's mid-layer states through the two maps — the only trained parameters — and scores the full library against compact per-skill banks built with a single forward pass at installation. A verdict then resumes the shortlisted skills' forward passes and reads the model's own likelihood and yes/no judgment, fused with the glance score as a product of experts.
According to the authors, the backbone stays frozen and the two projections contain just 7.9M trainable parameters on Qwen3-32B. Each new skill is indexed with one frozen forward pass, with no per-skill training. Trained once, Gavel is reported to transfer zero-shot to three public benchmarks and to SkillTraj, a new benchmark of 372 simulated agent trajectories that tests routing when the need for a skill arises mid-rollout.
On Qwen3-32B, the paper reports gains of up to 13.4 points on written tasks and up to 21.9 points when the need for a skill arises mid-rollout, compared with progressive disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters. Routing accuracy is said to improve with the backbone, and in a bash-agent harness the same 32B model reportedly triggers the correct skill on Skill-Use more often than larger frontier models running in Codex.