The Router Within: Eliciting Native Skill Routing from a Frozen LLM
A frozen LLM can expose its own skill-routing signal with only two trained linear maps.
The paper introduces Gavel, which scores a skill library without putting skill text into the model context. It builds compact per-skill banks at installation, then combines a fast “glance” shortlist with a “verdict” from resumed forward passes. The authors report zero-shot transfer across three public benchmarks and a new 372-trajectory SkillTraj benchmark. On Qwen3-32B, they say it beats progressive disclosure and retrieve-and-rerank systems, especially when a needed skill appears mid-rollout. ArXiv · AI/CL/LG's note
The paper introduces Gavel, which scores a skill library without putting skill text into the model context. It builds compact per-skill banks at installation, then combines a fast “glance” shortlist with a “verdict” from resumed forward passes. The authors report zero-shot transfer across three public benchmarks and a new 372-trajectory SkillTraj benchmark. On Qwen3-32B, they say it beats progressive disclosure and retrieve-and-rerank systems, especially when a needed skill appears mid-rollout. ArXiv · AI/CL/LG's note
score 5