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

Forum topic · 小凯 · 2026-09-16

Summary

Gavel (Glance And Verdict) is a skill-routing method showing that a frozen agent LLM already carries its own routing signal within its forward passes, readable via two trained linear maps without any skill text in the context. A 'glance' step projects the task's and each skill's mid-layer states to score the full skill library against compact per-skill banks built with one forward pass at install time; a 'verdict' step restores shortlisted skills' forward passes, reads the model's own likelihood 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 plus the new SkillTraj 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 in a bash agent harness the same 32B model triggered correct skills on Skill-Use more often than larger frontier models running in Codex. arXiv: 2609.15982.

Paper Overview

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

English Translation of the Abstract

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. A glance projects the task's and each skill's mid-layer states through the two maps — the only parameters trained — and scores the full library against compact per-skill banks built with one forward pass at install time. A verdict restores the forward passes of shortlisted skills, reads the model's own likelihood and yes/no judgments, and fuses them with the glance scores via a product of experts.

Gavel is trained once and transfers zero-shot to three public benchmarks plus 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 points when skill needs emerge mid-rollout. Routing accuracy scales with backbone capability; in a bash agent harness, the same 32B model triggers the correct skill on Skill-Use more often than larger frontier models running in Codex.

Key Points

Frozen LLMs contain a native routing signal in their forward passes; two linear maps are enough to extract it.
Only the two linear maps are trained — the backbone stays frozen.
Outperforms pipelines adding 1.2B–16B external parameters by up to 13.4–21.9 points.
Zero-shot transfer to three public benchmarks and the new SkillTraj benchmark (372 trajectories).