simonlqy/SkillGate-9B

SkillGate-9B

Policy from "SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents."

Agent frameworks expose skills — instruction files with a name, a one-line description and a body — by progressive disclosure: the agent sees only names and descriptions and must open a file to learn what is inside. With thousands of skills in a library, which one to read becomes a decision the policy makes mid-episode, and outcome-rewarded RL cannot teach it: the tokens naming the chosen skill carry a median 0.14% of their trajectory's loss weight, and two in five of them receive a negative advantage because execution afterwards failed.

SkillGate partitions one trajectory's token support into two disjoint credit channels: outcome credit reaches only execution tokens (the whole skill-read call is removed from the task loss), while an action-local advantage reaches exactly the skill-naming tokens, positive only when the trajectory's single read is the correct skill.

Model

Results (385-trial protocol, 5 agentic benchmarks, 16-candidate slate)

Same initialisation, data, steps and hyperparameters as the outcome-only row; the only difference is which tokens the gradient reaches.

Intended use

Research on agentic skill/tool selection. The model expects the OpenClaw-style prompt profile and tool schema used in the paper; see the repository for the exact system prompt and the frozen skill slates.

Links

• —Paper: arXiv:2608.18852
• —Code: https://github.com/DeepExperience/SkillGate

License

Derived from Qwen3.5-9B and distributed under the Qwen license; see license_link.