A self-play loop that grows its own skill library, not just harder tasks

Skill Self-Play (arXiv, submitted 24 July 2026, Siyuan Huang and 12 co-authors from Alibaba’s Qwen group) targets the standing trade-off in LLM self-evolution between task diversity and verification reliability. Three components co-evolve in one loop: a proposer that generates tasks conditioned on dynamically sampled skills, a solver that explores candidate solutions, and a skill controller that reads execution feedback to update and expand the skill library. The authors report the method "consistently pushes the performance ceiling of competent backbones" and produces large turnarounds on initially misaligned models across tool-use and reasoning benchmarks; the abstract carries no figures, so the size of the effect is not yet checkable from the listing.

What it means A persistent, growing skill library is the piece most agent stacks lack — every run starts cold — so the mechanism is worth tracking even before the numbers are public.