SkillScaffold: Scaffolding Agentic Reinforcement Learning with Policy-Adaptive Skill Credit Assignment and Selective Injection
SkillScaffold decides when a reusable “skill” should guide an agent, and when the agent should proceed without one.
The paper proposes a training scaffold for agentic RL that estimates skill utility from outcome-verified trajectories instead of running extra rollouts for each candidate. It uses likelihood comparisons to choose useful skills and abstains when none are expected to help. The authors report gains over agentic-RL and skill-based RL baselines on ALFWorld, WebShop, ScienceWorld, and τ²-Bench. ArXiv · AI/CL/LG's note
The paper proposes a training scaffold for agentic RL that estimates skill utility from outcome-verified trajectories instead of running extra rollouts for each candidate. It uses likelihood comparisons to choose useful skills and abstains when none are expected to help. The authors report gains over agentic-RL and skill-based RL baselines on ALFWorld, WebShop, ScienceWorld, and τ²-Bench. ArXiv · AI/CL/LG's note
score 5