When letting a model generate its own problems, solve them, and train on them, the biggest fear is that incorrect problems get absorbed into learning as well.

In this paper, the Qwen team proposes Skill Self-Play, adding a continuously updated skill library to the model’s self-evolution process.

Three roles are involved in training:

The Proposer generates tasks that are just challenging enough based on current skills, the Solver attempts to solve them, and the Controller updates the skill library based on execution results: effective skills are further refined, outdated skills are eliminated, and new patterns are distilled into new skills.

The system also retains an open-ended exploration path, preventing the model from repeatedly generating tasks around existing skills and gradually losing task diversity.

Here, skills evolve from an inference-time plug-in into a curriculum and verification interface during training. They both constrain task structure and help the system judge whether answers are reliable.

Experiments cover 5 open models ranging from 3B to 14B. Tool-calling capability improved by up to 42.9 percentage points, and logical reasoning by up to 12 percentage points.

More interestingly, the final model does not read these skill packs during either training or evaluation. Skills are responsible for organizing the learning process, and the learned capabilities ultimately become part of the model itself.

What I find important about this work is that it adds a maintainable intermediate structure to “model self-evolution.”

Future model training may gradually shift from manually preparing fixed data toward the co-evolution of tasks, capabilities, and curricula.

arxiv: https://arxiv.org/abs/2607.22529

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

Source: https://arxiv.org/abs/2607.22529 Authors: Siyuan Huang (https://arxiv.org/search/cs?searchtype=author&query=Huang,+S), Pengyu Cheng (https://arxiv.org/search/cs?searchtype=author&query=Cheng,+P), Haotian Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+H), Tao Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+T), Yihao Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+Y), Jingwei Ni (https://arxiv.org/search/cs?searchtype=author&query=Ni,+J), Shijie Zhou (https://arxiv.org/search/cs?searchtype=author&query=Zhou,+S), Ziyi Yang (https://arxiv.org/search/cs?searchtype=author&query=Yang,+Z), Gangwei Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+G), Mengyu Zhou (https://arxiv.org/search/cs?searchtype=author&query=Zhou,+M), Yu Cheng (https://arxiv.org/search/cs?searchtype=author&query=Cheng,+Y), Xiaoxi Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+X), Guanjun Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+G)

View PDF (https://arxiv.org/pdf/2607.22529)

Abstract: LLM training is shifting from manual design and annotation to interaction-driven self-evolution. However, existing self-evolutionary methods face a fundamental dilemma between task diversity and verification reliability: environment-bound methods obtain precise feedback but confine learning to narrow domains, while open-ended self-generation broadens the task space but lacks reliable verification, allowing misleading rewards to pollute the training loop. We identify agent skills as a powerful middle ground to reconcile this tension: each skill ensures deep, verifiable execution in a specific scenario, while dynamic routing across skills maintains open-ended task variety. Leveraging this insight, we introduce Skill Self-Play (Skill-SP), a co-evolutionary framework comprising a proposer, a solver, and a dynamic skill controller. Orchestrated via a reinforcement learning loop, these components co-evolve in a continuous self-play loop: the proposer generates challenging tasks conditioned on dynamically sampled skills; the solver explores candidate solutions to push its capability boundaries; and the skill controller collects execution feedback to update and expand the skill library. This interactive co-evolution effectively bridges the gap between structured verification and open-ended exploration. Empirical evaluations on tool-use and reasoning benchmarks demonstrate that Skill-SP, serving as a robust evolution engine, consistently pushes the performance ceiling of competent backbones while catalyzing striking turnarounds for initially misaligned models. Our code is available at this https URL (https://github.com/Qwen-Applications/skill-self-play).

Submission history

From: Siyuan Huang [view email (https://arxiv.org/show-email/7c215241/2607.22529)] [v1] Fri, 24 Jul 2026 17:59:22 UTC (6,994 KB)