Skill1 framework trains AI agents to co-evolve skill selection, utilization, and distillation from a single task-outcome reward signal

Skill1 is a framework that trains a single policy to handle three coupled capabilities: the agent queries and re-ranks a skill library to select a relevant skill, uses it to solve a task, and distills new skills from its experience trajectory.

All learning derives from a single task-outcome signal, with its low-frequency trend crediting skill selection and high-frequency variation crediting skill distillation, allowing the three capabilities to co-evolve toward a shared objective rather than optimizing in isolation.

Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and reinforcement learning baselines; ablations confirm that removing any credit signal degrades the evolution of the three capabilities.