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

Researchers introduced Skill Self-Play, a co-evolutionary RL framework for LLMs with a proposer, solver, and dynamic skill controller.

It uses agent skills as a middle ground between narrow environment-based training and unreliable open-ended self-generated tasks.

The approach improved performance on tool-use and reasoning benchmarks, including Qwen-Applications.

It offers a more scalable and reliable path for LLM self-improvement without depending only on hand-designed data.