simonlqy/SkillGate-9B
The simonlqy/SkillGate-9B model is a 9 billion parameter language model based on Qwen3.5-9B, specifically fine-tuned for improved skill selection in long-horizon agentic tasks. It utilizes a novel credit assignment mechanism to enhance the policy's ability to choose the correct skill from a large library. This model is primarily intended for research in agentic skill and tool selection, demonstrating superior performance in identifying relevant skills compared to traditional SFT and outcome-rewarded RL methods.
Loading preview...
SkillGate-9B: Enhanced Skill Selection for AI Agents
SkillGate-9B is a 9 billion parameter model, built upon the Qwen3.5-9B architecture, designed to address the challenge of effective skill selection in complex, long-horizon AI agent tasks. Traditional reinforcement learning often struggles with credit assignment for skill selection due to the delayed and diluted impact of skill-naming tokens on the overall task loss.
Key Capabilities & Innovations
- Novel Credit Assignment: SkillGate introduces a partitioned credit system, where outcome credit is directed only to execution tokens, and a distinct action-local advantage specifically rewards skill-naming tokens when the correct skill is chosen. This mechanism significantly improves the policy's ability to learn effective skill selection.
- Superior Skill Selection Performance: In a 385-trial protocol across 5 agentic benchmarks, SkillGate-9B achieved an overall success rate of 53.2%, markedly outperforming SFT (40.8%) and outcome-only SkillRL (47.0%). It demonstrated exceptional accuracy in "Oracle read" scenarios (83.9%) and significantly reduced "Misleading read" instances (21.8%).
- Targeted Training: The model was trained for 100 steps using on-policy GRPO, focusing on 491 tasks with specific hyperparameters to optimize skill selection.
Intended Use Cases
- Research in Agentic AI: This model is specifically developed for research into agentic skill and tool selection, particularly within frameworks that expose skills through progressive disclosure.
- OpenClaw-style Prompting: It expects an OpenClaw-style prompt profile and tool schema, aligning with the methodology presented in the associated research paper.
For more technical details, refer to the SkillGate paper and the code repository.