Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

TL;DR - Skill Self-Play is a reinforcement-learning framework in which task generation, problem solving, and an agent skill library co-evolve. It aims to combine open-ended task diversity with reliable, skill-based verification.

• A proposer generates challenging tasks conditioned on dynamically sampled skills.
• A solver explores candidate solutions, while a controller uses execution feedback to update and expand the skill library.
• Dynamic skill routing supports broad exploration while preserving scenario-specific, verifiable execution.
• Evaluations on tool-use and reasoning benchmarks report improvements for capable backbones and initially misaligned models.