Researchers introduce Gavel, a skill-routing method that extracts selection signals directly from a frozen LLM’s mid-layer activations using only two trained linear projection maps, avoiding the need to preload skill metadata into context or offload routing to external retrieval systems. A two-stage ‘glance and verdict’ process scores the full skill library via cached per-skill representations, then reruns forward passes on shortlisted skills to read the model’s own confidence judgment. Trained once, Gavel transfers zero-shot to three public benchmarks and a new 372-trajectory benchmark (SkillTraj), outperforming larger baselines by up to 21.9 points while letting skill libraries scale past context-window limits.
