MSR-Orchard / WildAWM

Agent World Model in the Wild 
 Computer-Use Agent Environment Synthesis 
 Agent World Model in the Wild (WildAWM) synthesizes executable personal-computer worlds and grounded cross-interface (XAI) tasks for general computer-use agents. Each world combines local files and command-line workflows, MCP services, and mirrored websites with graphical interfaces from WebHarbor , and evolves day by day from a persona. Each task is validated for solvability, difficulty, and verifier reliability through ensemble execution. The entire pipeline is powered by open-source LLMs, without any proprietary models . 
---
## 📣 News
- **Oct 19, 2026** 📢 WildAWM paper, code, environments and tasks, and trained checkpoints released.
## 🔮 Resources
| Resource | Link |
|---|---|
| 📄 Paper | arXiv: 2610.26828 |
| 🌐 Project page | msr-orchard.github.io/WildAWM |
| 💻 Code | MSR-Orchard/WildAWM |
| 🤗 Data and models | Hugging Face collection |
| 📦 341 environment-task pairs | ChilleD/WildAWM |
| 🧭 Harbor hub | wildawm |
| 🤖 WildAWM-35B-A3B | ChilleD/WildAWM-35B-A3B |
| 🤖 WildAWM-4B | ChilleD/WildAWM-4B |
## 📦 Installation
Requires Python 3.12+, uv, Docker, and Node.js. One `uv sync` installs both synthesis and evaluation, and all commands below run from the repository root.
 [code] 
## 🔥 Synthesis
Synthesis has two stages. **World synthesis** builds cumulative personal-computer states $W(c,d)$ from a persona $c$ and a day $d$. **XAI task synthesis** explores these states, authors tasks and verifiers, collects heterogeneous agent rollouts, and applies five quality gates. Accepted `{task, environment, verifier}` triplets support supervised fine-tuning, online reinforcement learning, and evaluation.
To directly use the released 341 environment-task pairs instead, download them and go straight to evaluation:
 [code] 
### Setup
Copy the two configuration files, fill in your endpoints, keys, and model refs, and export `.env` into the shell that runs synthesis commands:
 [code] 
Build the runtime images, then copy the pristine seeds of the 15 WebHarbor website mirrors from the pinned WebHarbor image (once per data root):
 [code] 
 Main settings in synthesis/.env and synthesis/models.json 
| Setting | Purpose |
|---|---|
| `models.json` | Providers, endpoints, keys, model ids, context windows, and reasoning-effort support. Models are referenced as `provider/model`. |
| `WILDAWM_SYNTH_MODEL_REF` | Default generator and judge model. |
| `WILDAWM_VISION_BASE_URL`, `_API_KEY`, `_MODEL` | Vision model behind `describe_image` for text-only synthesis models. |
| `WILDAWM_ENSEMBLE_MODELS`, `WILDAWM_ENSEMBLE_HARNESSES` | Model × harness pool for the XAI rollout gates. Harnesses: `opencode`, `codex`, `claude-code`, `nanobot`, `openclaw`, `hermes`. |
| `WILDAWM_DATA_ROOT` | Optional. Moves generated worlds, tasks, outputs, and logs out of `synthesis/`. |
| `WILDAWM_SYNTHESIS_CONFIG` | Optional. Points at a copy of `synthesis/configs/synthesis.example.jsonc` with iteration limits and quality-ga