Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning (arxiv.org)

A groundbreaking framework named Skill1 has been introduced for the evolution of skill-augmented agents using reinforcement learning. This framework addresses a critical limitation in existing methods, which typically optimize skill selection, utilization, and distillation in isolation. Instead, Skill1 unifies these capabilities under a single policy, allowing agents to leverage a persistent skill library effectively. It enables agents to select and utilize relevant skills from the library while concurrently distilling new skills from their experiences—all guided by a shared task-outcome objective. This holistic approach enhances the agent's performance across varied tasks. The significance of Skill1 lies in its potential to advance skill-based learning by streamlining the co-evolution of multiple capabilities. Experiments conducted on the ALFWorld and WebShop environments demonstrate that Skill1 significantly outperforms previous skill-based and reinforcement learning models. The training dynamics confirm that the co-evolution of skill selection, execution, and distillation is crucial for achieving superior task outcomes. By employing a single task-outcome signal to guide the learning process, Skill1 not only improves efficiency but also ensures that each learning strategy contributes meaningfully to the agent's overall capability development.