The authors propose Skill Self-Play (Skill-SP), a co-evolutionary framework for LLM training that resolves a central tension: environment-bound methods give reliable feedback but narrow task scope, while open-ended generation broadens tasks but risks misleading rewards. Skill-SP uses as a middle ground, ensuring verifiable execution within each skill while maintaining variety through dynamic skill routing. The framework consists of a proposer that generates tasks conditioned on sampled skills, a solver that explores solutions, and a skill controller that updates the skill library based on execution feedback, all orchestrated via .
Empirical evaluations on tool-use and reasoning benchmarks show that Skill-SP consistently pushes the performance ceiling of capable backbone models and even turns around initially misaligned models. The results demonstrate that the co-evolutionary self-play loop effectively bridges structured verification with open-ended exploration, leading to broad improvements across both tool calling and logical reasoning tasks.