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

Source: [arXiv](https://arxiv.org/abs/2609.15982v1)  
Feed7 permalink: https://feed7.dev/p/2609-15982v1-1b049zj  
Published: 2026-09-14T17:58:27.000Z  
Trust: Needs Review (needs_review)

## Why Included

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.

## Source Summary

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.

## Practical Implication

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**.

## Agent-Ready Context

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.

## Connected Context

Feed7 judgment across 778 accumulated Signals:

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.

- [Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSet](https://feed7.dev/p/skills-are-new-features-building-skill-centric-harness-yogendra-miraje-f-0lp4c7o) — Gavel offers a concrete routing mechanism for the skill-centric harness, while FactSet’s account supplies the surrounding requirements—versioning, model-specific evaluation, access control, and governance—that routing accuracy alone does not cover.
- [500 Skills, Zero Fine-Tuning: LinkedIn's Playbook for AI Agents — Ajay Prakash, LinkedIn](https://feed7.dev/p/500-skills-zero-fine-tuning-linkedin-s-playbook-for-ai-agents-ajay-praka-1duois1) — LinkedIn demonstrates progressive discovery through search and meta-tools at large catalog scale; Gavel proposes a contrasting native-model shortlist-and-verdict pipeline that keeps skill descriptions out of active context.
- [Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind](https://feed7.dev/p/agents-without-code-skills-yaml-and-filesystems-replaced-python-philipp-0t7a4s8) — The low-code agent pattern reduces bespoke orchestration but leaves evaluation to builders; Gavel makes routing evaluation especially necessary because skill selection is learned separately from execution.
- [zhaoxuya520/reverse-skill](https://feed7.dev/p/reverse-skill-1e4jlfw) — The security router shows why selection must be paired with authorization and evidence stages; Gavel may improve which skill is chosen, but its reported design does not replace external controls around risky execution.

## Context Map

- Layer: agent
- Domains: coding
- Topics: skills, tool-use, context-engineering

## Uncertainty

- 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.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.