Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills
Alibaba
The Qwen team at Alibaba proposed Skill-SP, a reinforcement-learning framework with three co-evolving components: a proposer that generates tasks conditioned on dynamically sampled skills, a solver that explores solutions, and a skill controller that updates a skill library from execution feedback. It aims to bridge reliable-but-narrow environment-bound training with broad-but-unreliable open-ended task generation, showing gains on tool-use and reasoning benchmarks, including large improvements for initially weak models.
Why it matters
Ranked #2 on Hugging Face Daily Papers with 38 upvotes; addresses a real tension in current agent-training pipelines between verifiable narrow tasks and diverse but noisy self-generated ones.
Importance: 2/5
Notable RL/agentic training research; ranked #2 on HF Daily Papers.