We tested the paper’s central tool-call claim: whether skill-conditioned self-play generates more valid, frontier-targeted tasks than unguided self-play and produces a larger held-out solver gain. In a matched, two-seed, one-iteration reproduction, Skill-SP gained +1.46 seven-subset accuracy points versus +0.96 for unguided (paper, five iterations: +6.5 versus +3.9), but its structural, contract-valid, and frontier yields were lower by 0.29, 5.47, and 1.51 points. Assessment: partially aligned—the downstream direction matched, while the requested mechanism-yield claim did not under this bounded setup.
The reproduction pinned the released code and Qwen3-4B-Instruct checkpoint, used 1,024 proposals and five eight-group solver updates per arm/seed, and evaluated API-Bank Levels 1–3 plus BFCL JavaScript, Python, Java, and Live at avg@2. This is downscaled from five iterations and 8,000 proposals per iteration. Runs used Kubernetes on NVIDIA RTX PRO 6000 Blackwell GPUs, with 16 GPUs peak concurrent and 2.209151 observed campaign hours.
Read the illustrated report · Explore the self-contained notebook · Download exact measurements
| Branch / experiment | Purpose | Exact run command | Assessment / outcome | Compute |
|---|---|---|---|---|
main |
Public report and reproduction harness | Not run as an experiment (publication surface) | Presentation-only | — |
| Skill-SP seed 1 | Skill-conditioned arm | bash repro/run.sh |
+0.88 accuracy points; 15.1% frontier yield | Kubernetes, 8× RTX PRO 6000 Blackwell, 21m28s |
| Unguided seed 1 | Matched control | bash repro/run.sh |
+1.15 accuracy points; 16.7% frontier yield | Kubernetes, 8× RTX PRO 6000 Blackwell, 19m38s |
| Skill-SP seed 2 | Stochastic replicate | bash repro/run.sh |
+2.03 accuracy points; 18.0% frontier yield | Kubernetes, 8× RTX PRO 6000 Blackwell, 21m32s |
| Unguided seed 2 | Paired stochastic control | bash repro/run.sh |
+0.77 accuracy points; 19.4% frontier yield | Kubernetes, 8× RTX PRO 6000 Blackwell, 19m44s |
Self-play agent training must balance reliable supervision with evolving task coverage. Fixed tasks are verifiable but cannot follow a solver's changing capabilities, while unconstrained generation can introduce invalid or uninformative samples. Skill Self-Play (Skill-SP) closes this loop through skill-routed task generation, automatic validity verification, frontier-based curriculum construction, and feedback-driven skill evolution. It supports tool-call prediction and logical reasoning; skills are used only during training, while the final solver remains prompt-only at inference.
- Evolving skill library: modular skill packages provide task-generation guidance, and a router selects skills for each proposal.
- Verified task generation: the proposer generates tasks conditioned on the selected skill. Validity checks enforce task contracts, while the current solver empirically estimates each task's difficulty and utility.
- Frontier-based co-evolution: tasks near the solver's learning frontier form the training pool. The solver learns from this pool, while the proposer and solver are optimized for their respective objectives.
- Targeted skill evolution: the controller aggregates check failures, novel samples, and success and utility signals to trigger skill refinement, pruning, and induction; the resulting packages update the skill library.
Results reported in the paper. Tool-call overall is the combined API-Bank/BFCL average; logical overall is ZebraLogic grid-level accuracy.
| Backbone | Tool-call | Gain | Logical | Gain |
|---|---|---|---|---|
| Qwen3-4B-2507 | 60.2 -> 66.7 | +6.5 | 72.1 -> 73.5 | +1.4 |
| Qwen3-8B | 69.4 -> 72.2 | +2.8 | 23.6 -> 32.4 | +8.8 |
| Ministral-3-8B | 20.7 -> 63.6 | +42.9 | 5.0 -> 11.2 | +6.2 |
| Ministral-3-14B | 22.2 -> 64.5 | +42.3 | 5.4 -> 17.4 | +12.0 |
| Granite-4.1-3B | 57.2 -> 62.5 | +5.3 | 11.6 -> 12.6 | +1.0 |
Python 3.11 is recommended.
python -m pip install -U pip
python -m pip install -r requirements.txtInstall CUDA-compatible PyTorch, vLLM, and FlashAttention builds first when your
environment requires custom wheels. Set HF_TOKEN when gated Hugging Face
resources require authentication. For experiment logging, set WANDB_API_KEY
or use WANDB_MODE=offline.
API-Bank, BFCL, and ZebraLogic evaluation data are bundled under benchmarks/,
so the provided launchers run without additional dataset setup.
Set the model and output roots, then run one launcher with bash <script>.
The released launchers assume 8 visible GPUs: the questioner uses GPUs 0-3,
and the solver uses all 8 GPUs. Each launcher evaluates the base model and every
solver checkpoint automatically.
export SKILL_SP_OUTPUT_DIR=/path/to/output
export SKILL_SP_MODEL_ROOT=/path/to/model/checkpoints
bash train-tool-call-skill-qwen3-4b.sh| Backbone | Tool-call prediction | Logical reasoning |
|---|---|---|
Qwen3-4B-Instruct-2507 |
train-tool-call-skill-qwen3-4b.sh |
train-logical-skill-qwen3-4b.sh |
Qwen3-8B |
train-tool-call-skill-qwen3-8b.sh |
train-logical-skill-qwen3-8b.sh |
Ministral-3-8B-Instruct-2512-BF16 |
train-tool-call-skill-ministral3-8b-bf16.sh |
train-logical-skill-ministral3-8b-bf16.sh |
Ministral-3-14B-Instruct-2512-BF16 |
train-tool-call-skill-ministral3-14b-bf16.sh |
train-logical-skill-ministral3-14b-bf16.sh |
granite-4.1-3b |
train-tool-call-skill-granite-4.1-3b.sh |
train-logical-skill-granite-4.1-3b.sh |
We thank the contributors to VERL and R-Zero for their helpful open-source contributions to this work.
Apache License 2.0. See LICENSE.
If you find Skill-SP useful, please consider citing:
@article{huang2026skill,
title={Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills},
author={Huang, Siyuan and Cheng, Pengyu and Liu, Haotian and Chen, Tao and Liu, Yihao and Ni, Jingwei and Zhou, Shijie and Yang, Ziyi and Jiang, Gangwei and Zhou, Mengyu and Cheng, Yu and Jiang, Xiaoxi and Jiang, Guanjun},
journal={arXiv preprint arXiv:2607.22529},
year={2026},
url={https://arxiv.org/abs/2607.22529}
}