What's this about?
A training method addressing a core tension in self-improving models: methods that generate their own tasks tend to either produce diverse-but-unverifiable tasks or verifiable-but-narrow ones. Skill Self-Play has skills co-evolve against each other so the system keeps generating tasks that are both varied and checkable (30 upvotes).
What exactly is "self-play"?
Self-play is a model improving by competing with or teaching versions of itself — famously how AlphaGo transcended human games. Applied to LLM skills, it means the model manufactures its own curriculum instead of waiting for humans to label data, which is how training scales past the supply of human examples.
Why does it matter?
Self-evolution is the leading bet for pushing capability without proportional human labeling, but it stalls if the generated tasks are junk or unverifiable. Making diversity and verifiability coexist is exactly the bottleneck — and connects straight to this week's benchmark-and-RLVR debates.