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

Forum topic · 小凯 · 2026-09-16

Summary

This paper introduces Gavel (Glance And Verdict), a skill-routing method for LLM agents that reads routing signals directly from a frozen agent LLM's internal states instead of relying on in-context skill metadata or external retrieval pipelines. The authors show that a frozen LLM already carries the routing signal in its own forward passes, and that just two linear maps are enough to extract it with 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 trained linear maps, scoring the full library against compact per-skill banks built from a single forward pass at installation time; a 'verdict' restores the forward passes of shortlisted skills, reads the model's own likelihoods and yes/no judgments, and fuses them with glance scores via a product of experts. Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, a new benchmark of 372 simulated agent trajectories. On Qwen3-32B, it outperforms progressive disclosure and retrieval-rerank pipelines (which add 1.2B–16B external parameters) by up to 13.4 points on text tasks and up to 21.9 points when skill needs emerge mid-rollout. Routing accuracy scales with backbone capability, and a 32B model in a bash agent harness triggers correct skills more often than larger frontier models running in Codex on Skill-Use.

Paper Overview

Field: NLP
Authors: Ruishuo Chen, Xun Wang, Yu Chen, Zhuoran Li, Longbo Huang
Published: 2026-09-14
arXiv: 2609.15982

Key Points

Skills extend LLM agents beyond their parametric knowledge, but the benefit depends on choosing the right skill.
Current deployed harnesses route by preloading each skill's metadata into context, which disperses the agent's attention and caps the library size. Retrieval pipelines move selection out of the context — but also out of the agent's own capability.
Core finding: a frozen agent LLM already carries the routing signal in its own forward passes, and two linear maps suffice to read it out with no skill text in the context.

How Gavel Works

Gavel (Glance And Verdict from a frozen LLM) reads the native routing signal in two steps:

1. Glance: projects the task's and each skill's mid-layer states through two linear maps (the only trained parameters), scoring the entire library against compact per-skill banks built with a single forward pass at installation time. 2. Verdict: restores the forward passes of shortlisted skills, reads the model's own likelihoods and yes/no judgments, and fuses them with the glance scores via a product of experts.

Results

Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, a new benchmark of 372 simulated agent trajectories.
On Qwen3-32B, it beats progressive-disclosure and retrieval-rerank pipelines (which add 1.2B to 16B external parameters) by up to 13.4 percentage points on text tasks, and by up to 21.9 percentage points when skill requirements appear mid-rollout.
Routing accuracy improves with backbone capability: the same 32B model in a bash agent harness triggers the correct skill more often on Skill-Use than larger frontier models running in Codex.

Original Abstract (English)

> Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise rests on picking the right one. Deployed harnesses route by preloading every skill's metadata into the context, which disperses the agent's attention and caps the library size. Retrieval pipelines move the selection out of the context, but also out of the agent's capability. We show that the frozen agent LLM already carries the routing signal in its own forward passes, and that two linear maps suffice to read it out with no skill text in the context. Gavel (Glance And Verdict from a frozen LLM) reads it in two steps...

---

*Auto-collected on 2026-09-16.*