The Router Within: Eliciting Native Skill Routing from a Frozen LLM
Gavel routes skills from a frozen model’s internal states without loading skill text into context. It suggests large skill libraries may need model-native selection rather than metadata-heavy prompts.
Gavel reads skill-routing signals from a frozen LLM using **two trained linear maps**. A glance scores the full library from mid-layer states and prebuilt skill banks; a verdict resumes shortlisted forward passes and combines likelihood with a yes/no judgment.
Harness authors should benchmark routing separately from skill execution, especially when skills become relevant mid-run. This design keeps skill descriptions out of the active context and avoids external rerankers with **1.2B–16B parameters**.
Gavel reads skill-routing signals from a frozen LLM using **two trained linear maps**. A glance scores the full library from mid-layer states and prebuilt skill banks; a verdict resumes shortlisted forward passes and combines likelihood with a yes/no judgment. Harness authors should benchmark routing separately from skill execution, especially when skills become relevant mid-run. This design keeps skill descriptions out of the active context and avoids external rerankers with **1.2B–16B parameters**. On Qwen3-32B, Gavel leads compared pipelines by **up to 13.4 points** on written tasks and **up to 21.9 points** for mid-rollout skill needs. The claims span three public benchmarks plus a new simulated-trajectory benchmark, so production generality remains open.
Gavel shifts progressive skill discovery inside the frozen model: compact learned probes shortlist skills without placing the entire catalog in context or adding a large reranker, including when relevance appears mid-rollout. It strengthens the case for evaluating routing as its own subsystem, while leaving governance, execution safety, model-change reevaluation, and production transfer outside the reported routing gains.