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

Gavel reads native skill-routing signals from a frozen LLM's forward passes using two linear maps, beating retrieve-and-rerank pipelines that add 1.2B–16B external parameters by up to 21.9 points on Qwen3-32B.

Gavel (Glance And Verdict from a frozen LLM) shows that a frozen agent LLM already contains skill-routing signals in its forward passes, which can be read out by two trained linear maps without loading any skill text into context. A glance step scores the full skill library against compact per-skill banks built in a single forward pass at installation; a verdict step resumes shortlisted skills' forward passes and fuses the model's own likelihood and yes/no judgments as a product of experts. Trained once, it transfers zero-shot to three public benchmarks plus SkillTraj, a newly introduced benchmark of 372 simulated agent trajectories. On Qwen3-32B it beats progressive disclosure and retrieve-and-rerank pipelines that add 1.2B–16B external parameters by up to 13.4 points on written tasks and 21.9 points when skills are needed mid-rollout. Per the Hugging Face daily papers report, routing accuracy scales with backbone quality and Gavel beats larger models in the Codex harness. Both sources agree on all reported figures; no conflicts were found.

• Only two trained linear maps are needed; routing signal is read from the frozen LLM's mid-layer states with no skill text in context
• Glance step scores the full library using per-skill banks built in one forward pass at installation; verdict step fuses likelihood and yes/no judgments as a product of experts
• Zero-shot transfer to three public benchmarks plus SkillTraj, a new benchmark of 372 simulated agent trajectories
• On Qwen3-32B, beats progressive disclosure and retrieve-and-rerank pipelines adding 1.2B–16B external parameters by up to 13.4 points on written tasks and 21.9 points mid-rollout
• Routing accuracy scales with backbone quality; beats larger models in the Codex harness (per Hugging Face daily papers report)
• Sources: Hugging Face daily papers (2026-09-13) and arXiv cs.AI/cs.LG/cs.CL (2026-09-14); figures fully consistent across both reports