Skill Self-Play: Qwen Team Uses Agent Skills as the Middle Ground Between Verifiable Environments and Open-Ended Task Generation

arXiv 2607.22529·medium signalXBlueskyLinkedInCopy link

Skill-SP (arXiv 2607.22529, July 24, from the Qwen Business Unit) targets the core dilemma in self-evolving LLM training — environment-bound methods get precise feedback but stay narrow, while open-ended self-generation broadens tasks and lets misleading rewards pollute the loop. The insight is that agent skills reconcile both: each skill guarantees deep verifiable execution in a specific scenario while dynamic routing across skills preserves open-ended variety. The framework couples a proposer, a solver, and a dynamic skill controller in an RL loop where the proposer generates tasks conditioned on sampled skills, the solver explores solutions, and the controller collects execution feedback to update and expand the skill library. Evaluations on tool-use and reasoning benchmarks show it lifts the ceiling of already-competent backbones and produces the largest turnarounds on initially misaligned models.