The Router Within: Eliciting Native Skill Routing from a Frozen LLM
The paper claims a frozen LLM can route to skills from its own hidden states, without loading skill descriptions into context.
Gavel trains only two linear maps, then uses mid-layer activations to score a task against compact per-skill banks built at installation. A second “verdict” step resumes shortlisted skill forward passes and combines likelihood with a yes/no judgment. The authors report zero-shot transfer across three public benchmarks and their SkillTraj benchmark, with Qwen3-32B beating larger retrieve-and-rerank systems in the stated tests. HF Daily Papers' note
Gavel trains only two linear maps, then uses mid-layer activations to score a task against compact per-skill banks built at installation. A second “verdict” step resumes shortlisted skill forward passes and combines likelihood with a yes/no judgment. The authors report zero-shot transfer across three public benchmarks and their SkillTraj benchmark, with Qwen3-32B beating larger retrieve-and-rerank systems in the stated tests. HF Daily Papers' note
score 5