ChilleD/WildAWM

Agent World Model in the Wild

Computer-Use Agent Environment Synthesis

Agent World Model in the Wild (WildAWM) releases 341 accepted cross-interface (cli/mcp/web) tasks over frozen persona-day worlds in Harbor task format, split into 100 SFT training tasks, 208 RL training tasks, and 33 test tasks.

---

🔮 Resources

Resource	Link
📄 Paper	arXiv: 2610.26828
🌐 Project page	WildAWM
💻 Code	WildAWM repository
🤗 Data & models	Collection
🤖 WildAWM-35B-A3B	Model
🤖 WildAWM-4B	Model
🧭 Harbor hub	wildawm datasets
📦 100 SFT training tasks	dataset/train-sft-100-set
📦 208 RL training tasks	dataset/train-rl-208-set
📦 33 test tasks	dataset/val-33-set

📦 Setup

docker --version                 # Docker >= 20.10
uv tool install harbor           # Harbor CLI 0.23.x
cp env.example .env              # set OPENAI_* and WILDAWM_VISION_LLM_*
cp run.example.yaml run.yaml     # set agents[0].model_name

Configuration variables are documented in env.example. Agent, model, attempts, concurrency, and run-time plugins are in run.example.yaml.

🔥 Evaluation

Run harbor run --help for full command options.

# agent runs, configured in run.yaml (default: test split, 4 attempts, 8 concurrent)
harbor run -c run.yaml --env-file .env
harbor run -c run.yaml --env-file .env -p dataset/train-sft-100-set
harbor run -c run.yaml --env-file .env -p dataset/train-rl-208-set

# one task with an agent
harbor run -c run.yaml --env-file .env -p dataset/val-33-set/16bfe8-d8-h1

-c accepts a YAML JobConfig and CLI flags override it (-p replaces the dataset). Inspect the resolved config with harbor run -c run.yaml --print-config. Paths in the config are relative to the working directory, so run harbor from the repository root.

Verification is programmatic (tests/verify_wrapper.py): completed gives reward 1, failed gives reward 0, and world_imperfect or script_error writes no reward file, which Harbor records as a trial error.

Citation

If you find this work useful, please cite:

@article{wang2026agentworldmodelwild,
  title={Agent World Model in the Wild: Computer-Use Agent Environment Synthesis},
  author={Zhaoyang Wang and Qianhui Wu and Haorui Wang and Simon Yu and Baolin Peng and Canwen Xu and Hao Cheng and Wenlin Yao and Boyi Liu and Yite Wang and Jianfeng Gao and Huaxiu Yao},
  year={2026},
  eprint={2610.26828},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.26828},
}