PSN framework lets AI build and reuse skills like a growing neural network

Continual skill acquisition in open-ended environments remains a major challenge for AI agents. A new paper from Haochen Shi, Xingdi Yuan, and Bang Liu introduces the Programmatic Skill Network (PSN), a framework where skills are represented as executable symbolic programs that form a compositional network. The network evolves as the agent gains experience, mirroring the learning dynamics of neural network training. PSN leverages large language models (LLMs) to instantiate three core mechanisms: OpReflect for structured fault localization over skill compositions, progressive optimization with maturity-aware update gating (which stabilizes reliable skills while allowing plasticity for uncertain ones), and canonical structural refactoring under rollback validation to maintain network compactness.

The researchers tested PSN on MineDojo and Crafter, two challenging open-ended embodied environments. Results showed that PSN agents achieved robust skill reuse, rapid adaptation to new tasks, and strong generalization across diverse task distributions. The maturity-aware gating prevented catastrophic forgetting while enabling continuous learning, and the refactoring mechanism kept the skill library efficient. This work suggests that combining symbolic programmatic skills with LLM-driven evolution could be a powerful approach for building AI agents that learn and adapt continuously in the real world.