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

TL;DR

Skill-SP makes agent skills the unit that a self-play loop evolves, using each skill's narrow, verifiable execution to keep rewards honest while routing across a growing skill library keeps the task space open-ended. Trending on Hugging Face Daily Papers the week of July 27–28, 2026.

Authors & Org

Siyuan Huang, Pengyu Cheng, Haotian Liu, Tao Chen, Yihao Liu, Jingwei Ni, Shijie Zhou, Ziyi Yang, Gangwei Jiang, Mengyu Zhou, Yu Cheng, Xiaoxi Jiang, Guanjun Jiang. Affiliations: unknown — not stated in the listing captured (source).

Method

The paper names the dilemma it is built around: environment-bound self-evolution gets precise feedback but confines learning to narrow domains, while open-ended self-generation broadens the task space and loses reliable verification, letting misleading rewards pollute the training loop (arXiv).

Skill-SP's answer is a three-part co-evolutionary loop:

1. Proposer — generates challenging tasks conditioned on dynamically sampled skills
2. Solver — explores candidate solutions to push its capability boundary
3. Dynamic skill controller — collects execution feedback to update and expand the skill library

The claimed insight is that a skill is the right middle ground: it is narrow enough that execution inside it can be verified, and the library is broad enough that routing across skills keeps task variety open-ended.

Results

Evaluated on tool-use and reasoning benchmarks. The authors report that Skill-SP consistently pushes the performance ceiling of already-competent backbones, and produces large turnarounds for initially misaligned models (source).

Specific benchmark names and score deltas are not present in the listing captured; treat the results section as the authors' claims pending a full read.

Significance

• Third distinct attack on the same wall this month. AREX: Towards a Recursively Self-Improving Agent for Deep Research recursed on the research process, Self-Distilled Agentic Reinforcement Learning recycled the policy's own trajectories, and Skill-SP evolves the skill inventory. All three are trying to buy open-endedness without losing a verifiable reward.
• Skills as a first-class training object. Most agent-skill work treats skills as an inference-time artifact — something authored, retrieved and executed. Here the skill library is the thing being trained, which puts it on the same footing as the policy.
• "Initially misaligned models" is the interesting claim. If the turnaround result holds, the loop is doing more than sharpening a good model — it is repairing a bad one, which is a different and stronger assertion than most self-play results make.

Open Questions

• What stops skill-library drift? A controller that expands the library on execution feedback can also entrench a skill that merely looks verifiable.
• Verification inside a skill is only as good as the skill's environment. The paper's premise is that narrow scope buys reliable reward — unproven for skills whose execution is itself model-judged.
• No stated compute cost. A three-component co-evolving loop is expensive; without a budget comparison the ceiling gains are hard to weigh against simply training longer.
• Affiliation and code availability unknown as captured.

Cite

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

Related

• Agentic Reinforcement Learning — the training paradigm this sits inside
• Agents (LLM Agents) — skills are the agent-side abstraction being evolved
• MCP — Model Context Protocol — the protocol layer that makes skills portable across harnesses
• AREX: Towards a Recursively Self-Improving Agent for Deep Research — recursive self-improvement, same week