Dataset
Entry point for WIA-Corpus and WIA-Bench resources.
Hugging FaceLet digital agents understand the environment before taking action. WIA unifies seven digital worlds and uses a foundational world model for state prediction, simulated training, and proactive decisions.
Discover how WIA unifies world modeling and action learning across seven digital environments.
1080p · English narration with subtitles · 2 min 35 sec
Entry point for WIA-Corpus and WIA-Bench resources.
Hugging FaceEntry point for WIA-WM model weights and checkpoints.
Hugging FaceEntry point for training, evaluation, and deployment code.
GitHubWorld Intelligence Agent: Unifying World Modeling and Action Learning for Digital Agents
Technical ReportWIA is not a passive simulator; it extends environment prediction into full agent-policy optimization.
Full-lifecycle CPT, SFT, and RL training corpora that unify heterogeneous digital interactions into action–observation trajectories.
The first paired ID/OOD world-model benchmark with 1.8k in-domain and 1.7k out-of-domain samples across seven digital environments.
4B / 9B / 27B / 35B-A3B / 122B-A10B foundation world models for systematic state transitions, simulated SFT, simulated RL, and lookahead reasoning.
Couples prediction and execution through Dual-Perspective Inversion and CSPO so policies implicitly internalize environment dynamics.
Unified modeling across Android, MCP, OS, Search, SWE, Terminal, and Web, without siloing knowledge by interface.
From small models to the 122B-A10B MoE large model, validating that world-model capability scales with size for agent training and inference.
A unified model-family lineup covering efficient small models to large-scale MoE.
Efficient world-model baseline
Balanced reasoning and capability
High-performance dense tier
Efficient MoE scaling
Large-scale MoE
WIA covers the full loop from data infrastructure, world simulation, to action execution: WIA-Corpus and WIA-Bench establish the data/evaluation foundation, WIA-FWM models state transitions and reward dynamics, and WIA-WAM aligns action generation with predicted dynamics.
Serialize seven digital interfaces into a shared textual action–observation representation.
CPT → SFT → RL three-stage training builds systematic next-observation prediction.
Use the world model to generate simulated data, provide simulated rewards, and reason ahead.
DPI aligns dual-perspective representations, and CSPO directly links action choices with simulated outcomes.
WIA-WM achieves leading average world-model scores on WIA-Bench and significantly outperforms strong baselines on the OOD suite.
| Model | Android | MCP | OS | Search | SWE | Terminal | Web | Avg | MAvg |
|---|---|---|---|---|---|---|---|---|---|
| Proprietary | |||||||||
| GPT-5.6-luna | 48.37 | 44.51 | 54.96 | 40.01 | 55.88 | 62.42 | 53.87 | 51.91 | 51.43 |
| GPT-5.6-terra | 48.34 | 43.55 | 56.14 | 40.46 | 57.68 | 61.79 | 54.57 | 52.38 | 51.79 |
| GPT-5.6-sol | 49.39 | 45.75 | 59.70 | 39.99 | 58.61 | 63.63 | 55.00 | 53.58 | 53.15 |
| Claude-Sonnet-5 | 47.76 | 49.85 | 63.22 | 35.26 | 44.58 | 46.65 | 54.24 | 48.03 | 48.79 |
| Claude-Opus-4.8 | 48.98 | 53.97 | 63.02 | 39.10 | 58.15 | 57.93 | 55.26 | 53.40 | 53.77 |
| Open-source | |||||||||
| GLM-5.3 | 47.20 | 50.23 | 55.58 | 31.10 | 32.68 | 53.88 | 51.03 | 44.98 | 45.96 |
| Kimi-K3 | 50.50 | 38.74 | 63.05 | 41.25 | 56.67 | 62.75 | 53.48 | 53.19 | 52.35 |
| Qwen3.8-27B | 41.08 | 39.40 | 53.34 | 34.88 | 48.47 | 51.84 | 47.51 | 45.46 | 45.22 |
| Qwen3.8-max | 47.60 | 45.32 | 57.97 | 36.82 | 57.62 | 63.42 | 55.48 | 52.40 | 52.03 |
| DeepSeek-V4-Flash | 46.74 | 44.46 | 58.93 | 38.40 | 50.51 | 55.53 | 50.77 | 49.40 | 49.33 |
| DeepSeek-V4-Pro | 47.87 | 43.92 | 61.93 | 38.77 | 51.14 | 57.77 | 52.63 | 50.73 | 50.57 |
| Qwen3.5 | |||||||||
| Qwen3.5-4B | 39.92 | 38.38 | 55.29 | 30.44 | 37.26 | 39.50 | 45.24 | 40.46 | 40.86 |
| Qwen3.5-9B | 43.00 | 41.37 | 58.19 | 30.99 | 44.05 | 43.81 | 47.42 | 43.76 | 44.12 |
| Qwen3.5-27B | 47.84 | 42.56 | 64.29 | 33.75 | 48.10 | 55.09 | 52.78 | 49.15 | 49.20 |
| Qwen3.5-35B-A3B | 47.09 | 42.69 | 56.59 | 32.03 | 45.70 | 49.80 | 49.55 | 46.01 | 46.21 |
| Qwen-AgentWorld | 47.43 | 42.63 | 64.21 | 45.02 | 56.31 | 57.22 | 53.68 | 52.87 | 52.36 |
| Qwen3.5-122B-A10B | 43.92 | 32.43 | 41.86 | 34.93 | 48.65 | 55.70 | 44.15 | 44.07 | 43.09 |
| Ours | |||||||||
| WIA-WM-4B | 27.19 | 44.10 | 48.40 | 36.84 | 60.06 | 55.87 | 49.27 | 46.34 | 45.96 |
| WIA-WM-9B | 46.82 | 52.12 | 62.70 | 46.06 | 60.77 | 59.38 | 53.43 | 54.49 | 54.47 |
| WIA-WM-27B | 50.51 | 52.53 | 63.42 | 46.36 | 59.92 | 60.73 | 54.52 | 55.45 | 55.43 |
| WIA-WM-35B-A3B | 48.76 | 43.72 | 63.86 | 51.33 | 59.19 | 62.52 | 54.95 | 55.72 | 54.90 |
| WIA-WM-122B-A10B | 50.59 | 40.00 | 64.71 | 58.18 | 60.46 | 64.62 | 53.59 | 57.35 | 56.02 |
Red: highest score in each column. Blue: second-highest score. Tied scores share the same rank.
Watch demonstrations across seven digital environments and explore WIA data, models, and research.
Watch Demos