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

Researchers introduced Skill Self-Play, a reinforcement learning framework for large language model self-improvement.

It uses a task proposer, solver, and skill controller that co-evolve by sampling skills, generating tasks, solving them, and updating a skill library from execution feedback.

The approach aims to combine open-ended task diversity with reliable verification, avoiding the limits of narrow environments or weak self-generated tasks.

Tests on tool-use and reasoning benchmarks suggest the method can improve LLM capability and may help weaker models recover and scale better.