Researchers at Tsinghua University have built Gavel, a router that decides when an AI agent needs a skill package and which one to open. In an end-to-end test on 177 executable tasks, Gavel raised the correct-skill trigger rate of the same frozen Qwen3-32B model from 1.1% to 90.9%. The September 14 preprint addresses a practical failure: an agent may have the right instructions but fail to consult them when its work changes mid-task.
A skill packages instructions, scripts and reference files for a specific job. Progressive-disclosure systems place every installed skill's name and description in the prompt, leaving the model to open the appropriate file. That menu consumes more context as the library grows. Retrieval systems assign selection to separate embedding and reranking models, which do not share the working agent's internal representation of the task. The researchers tested whether that representation already contained a usable routing signal.
Gavel reads the signal in two stages. When a skill is installed, one pass through the frozen model turns its document into a compressed bank of intermediate-layer states. During a task, two trained linear projections compare the current token states with every bank and shortlist about nine candidates on average. A second pass reads each candidate with the current task, then combines the initial match, the task's likelihood given the skill and the model's yes-or-no relevance judgement. The backbone stays unchanged; the projections contain 7.9 million trained parameters.
The team trained the projections once on 51,104 queries covering 9,084 skills, then fixed them for three other libraries. Gavel beat the strongest comparison pipeline by 3.8 percentage points on SkillRet, 13.4 on SRA-Bench and 1.3 to 2.7 on Eval-Core. In 372 simulated multi-turn trajectories, where a skill became necessary after a user request, tool result, agent plan or failed first choice, its lead ranged from 8.6 to 21.9 points.
The Skill-Use test placed Gavel inside mini-swe-agent, a small software agent with terminal access, and used 177 tasks drawn from 79 skills. Its gate watched the model's output to decide when routing was needed; 17 loads occurred only after later command output entered the conversation. Gavel's 90.9% rate narrowly exceeded the 89.7% reported for an 8-billion-parameter embedder and an 8-billion-parameter reranker. The controlled comparison is the same Qwen3-32B agent with a prompt-based skill menu, which reached 1.1%.
| System in Skill-Use | Correct-skill trigger rate |
|---|---|
| Qwen3-32B with Gavel | 90.9% |
| Qwen3-Embedding-8B plus Qwen3-Reranker-8B | 89.7% |
| MiniMax-M3 with progressive disclosure in Codex | 86.4% |
| Qwen3-32B with progressive disclosure in mini-swe-agent | 1.1% |
The results come from an author-run preprint. SkillTraj's dialogues were generated by one language model and screened by another, while the written-task tests used model-based adjudication because their answer sets omit some valid skills. The end-to-end trial covered one harness and 79 curated skills. Gavel also needs hidden-state access, which ordinary cloud APIs do not expose; provider-side integration remains a proposal. Independent tests on changing enterprise libraries must measure routing accuracy, added latency and failures to abstain together.
Sources
- Ruishuo Chen and colleagues, “The Router Within: Eliciting Native Skill Routing from a Frozen LLM,” arXiv, September 14, 2026. Abstract, authorship, submission date and headline results.
- Full paper. Router design, training setup, benchmark sizes, comparison results, end-to-end trial, cost analysis and limitations. The comparison table was created by Global Tech News from Table 1. Lead photograph: Shixart1985 / Nenad Stojković, CC BY 2.0; the image is illustrative and does not show Gavel.
- Skill-Use benchmark paper. Benchmark scope and the progressive-disclosure results quoted by the Gavel study.
- mini-swe-agent repository. The open-source agent harness used for the end-to-end test.