The Router Within: Eliciting Native Skill Routing from a Frozen LLM
Gavel (Glance And Verdict from a frozen LLM) reads routing signals from a frozen LLM's forward passes using two linear maps, with no skill text in context. It projects task and skill mid-layer states through the maps, scores against per-skill banks built in one forward pass, then resumes shortlisted skills' forward passes to fuse the model's likelihood and yes/no judgment with the glance as a product of experts. Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, a new benchmark of 372 simulated agent trajectories. On Qwen3-32B it outperforms progressive disclosure and retrieve-and-rerank baselines.
A research paper introduces Gavel, a method for skill routing in LLM agents that extracts routing signals directly from a frozen model's forward passes using two trained linear maps. It avoids preloading all skill metadata into context and outperforms progressive disclosure and retrieve-and-rerank approaches on Qwen3-32B, transferring zero-shot to three public benchmarks and the new SkillTraj benchmark of 372 simulated agent trajectories.
Gavel uses a two-stage process: a glance stage projects mid-layer states of the task and each skill through two linear maps to score a full library against compact per-skill banks built in one forward pass; a verdict stage resumes forward passes for shortlisted skills and combines the model's own likelihood and yes/no judgment with the glance score via a product of experts. This demonstrates that frozen LLMs contain native routing signals that can be elicited with minimal additional parameters.
The approach could enable larger skill libraries for LLM agents without context dilution or external retrieval pipelines, potentially reducing latency and improving selection accuracy. It suggests a shift toward leveraging internal model representations for routing rather than relying on metadata preloading or separate retrieval systems.
For enterprises deploying LLM agents with many skills, Gavel could lower inference costs and improve reliability by avoiding context bloat and external retrieval overhead. It may enable more scalable agent architectures and reduce the need for fine-tuning or complex routing pipelines.
Next observable signals include open-source release of Gavel code or SkillTraj benchmark, adoption in agent frameworks, and extensions to other model families or multi-step routing. Further research may explore whether the linear maps generalize across tasks or require per-domain training.