# 阿里通义千问团队提出 Skill Self-Play 方法

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-07-28 11:05
- AIHOT 分数：36
- AIHOT 链接：https://aihot.virxact.com/items/cms43gus205q2ro82f8tq4nse
- 原文链接：https://x.com/rohanpaul_ai/status/2081939149758890249

## AI 摘要

阿里通义千问团队提出 Skill Self-Play 方法，通过构建不断增长的“技能库”指导 AI 生成高质量训练任务，解决自生成数据质量不可控的难题。该方法让模型利用已验证的任务模板生成更好的训练数据，而非在无监管的自生成数据上训练。

## 正文

Alibaba's Qwen Team's new paper. Skill Self-Play turns agent skills into a training-time curriculum engine.

Attacks a crucial problem： when an AI creates its own training tasks， it either stays inside narrow， easy-to-check environments or generates lots of bad and unreliable tasks.

Their solution is a growing "skill library" that tells the AI what kinds of tasks to create， how to create them correctly， and how to verify them.

These skills are used to generate better training data， not mainly to help the model answer questions.

So the whole idea is： let the model improve itself through an evolving collection of verified task templates， instead of training on uncontrolled self-generated data.

---

- arxiv. org/abs/2607.22529

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