Two linear maps beat context-stuffing at choosing an agent's skill
single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC
What this means for you
Nothing to install from what is described here — no code, checkpoint or licence is mentioned — and the gains are measured on one backbone family against the authors' own baselines. If your harness preloads every skill's description into the prompt, the claim worth checking is that library size stops mattering.
A paper reports that the signal for choosing which skill an agent should load already sits in a fixed model's internal states. Skills are reusable procedures a harness loads on demand. The method, Gavel, trains two linear maps — its only trained parameters — to score the whole library from middle-layer states of the task and each skill, with no skill text in the prompt; shortlisted skills then get a full forward pass and a yes/no judgement. The authors report it beats progressive-disclosure and retrieve-and-rerank baselines on Qwen3-32B, by up to 13.4 points on written tasks and 21.9 when the need for a skill appears mid-run. Evaluation dates are not given.
Models in this story
Key facts
- ·The paper describes Gavel, which trains two linear maps as its only trained parameters to route skills from a frozen model's mid-layer states, with no skill text in the prompt. source
- ·Reported gains on Qwen3-32B: up to 13.4 points on written tasks and up to 21.9 points when a skill is needed mid-rollout, against progressive-disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters. source
- ·The paper introduces SkillTraj, a benchmark of 372 simulated agent trajectories. source
- ·The paper reports that routing accuracy improves as the backbone model gets larger, and that in a bash-agent harness the 32B model triggers the correct skill on Skill-Use more often than larger frontier models running in Codex. source
- ·The paper was posted on 13 September 2026 and gives no evaluation dates for the reported benchmark scores. source
What the sources say
- Hugging Face Daily Papers (research) — Preprint describing a method that picks skills from a model's internal states, claiming gains over context-based selection.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersThe Router Within: Eliciting Native Skill Routing from a Frozen LLM2026-09-13