Qwen-AgentWorld: A Language World Model That Lets AI Rehearse Before It Acts
Imagine an AI agent playing an open-world game. A traditional LLM agent must actually connect to the server, open the webpage, run the command, and wait for the result—like a person who has to physically jump every time they wonder "what if I jumped from here?"
Qwen-AgentWorld builds a simulator inside the AI's head. When the agent asks "what happens if I grep this file?", the model can "imagine" the terminal output without executing it. When it asks "what happens if I click this button?", it predicts the next UI state.
This is the essence of a Language World Model (LWM)—not letting AI *perform* actions, but letting it *rehearse* their consequences.
Why This Matters: AI's Imagination
Humans have a core decision-making ability: simulating consequences before acting. Traditional AI agents lack this. Their loop is observe → think → act → wait for real feedback → think again. Every step requires real environment interaction—slow, costly, and irreversible. If the environment is a surgical robot or a trading system, the cost of trial and error is unbearable.
Qwen-AgentWorld's breakthrough: the language model itself becomes an interactive virtual environment. You can train agents inside it, letting them try various strategies—entirely within the model, no real servers, webpages, or phones needed.
Two Model Sizes
- 35B-A3B: 35B total parameters, 3B active (MoE). Runs on a single A100; deployable by SMEs.
- 397B-A17B: 397B total, 17B active (MoE). Peak performance, requires multi-GPU.
- Terminal: Linux command line (predicting cat/grep/pip output)
- SWE: software engineering (tracebacks, diffs)
- Search: search result pages
- MCP: API-returned JSON
- Web: DOM changes after clicks
- Android: UI hierarchy changes after touch
- Desktop OS: window states after mouse/keyboard input
- 9 points from a rubric judge: LLM-scored across 5 dimensions (format, factuality, consistency, authenticity, quality), 1–5 each.
- 1 point from a rule verifier: executable 0/1 checks as a hard anchor against reward hacking.
- Scalable: simulate thousands of environments concurrently
- Controllable: inject API errors, pagination, partial failures—rare edge cases agents must handle
- Better than real-world training: Sim RL beats Real RL by +4.3 to +7.1 on Claw-Eval/QwenClawBench (OpenClaw), +12.3 on MCPMark, +16.3 on WideSearch
- Across-the-board gains on 7 agent benchmarks
- RL trained only on Terminal data yielded +11.5 on SWE, +11.8 on Search, +5.0 on MCP—cross-domain emergence
- Paper: https://arxiv.org/abs/2606.24597
- Code & models: https://github.com/QwenLM/Qwen-AgentWorld
The dual sizing signals Qwen's bet that world models will move from lab to production line.
Three-Stage Training Pipeline
Stage 1: CPT — Seeing the World
Trained on 10M+ real environment interaction trajectories across 7 domains:
These trajectories were produced by real agents in real environments, not written by humans.
A clever engineering detail: an information-theoretic loss mask. Many turns in tool-call trajectories are "filler" (e.g., echo just returns its input). The team used four statistical metrics (overlap rate, novelty, Jaccard similarity, length ratio) to classify turns into 7 categories, keeping only 10%–50% of the loss for "filler" turns—preserving context without wasting compute on parroting.
Stage 2: SFT — Activating Next-State Prediction
SFT explicitly teaches the model "you are predicting the next state," producing long chain-of-thought reasoning: current state → what the action means → what the result should be given system logic → prediction.
SFT used rejection sampling: 3 rollouts per query, judge-scored, keeping only the top scores above a threshold. Out of 10,250 candidate queries, 7,094 survived (69.2% retention)—quality over quantity.
Stage 3: RL — Sharpening Simulation Fidelity
RL uses a hybrid reward:
The 9:1 ratio balances rich signal with a hard floor. The paper also notes a pitfall: expanding one trajectory into multiple turn-level samples creates shared long prefixes, causing reward collapse. The fix: expand each trajectory into only one turn, ensuring each sample has an independent prediction target.
AgentWorldBench: Evaluation
Traditional metrics (BLEU, ROUGE) only measure literal similarity—world model outputs may say the same thing differently, or include non-reproducible details like IDs and timestamps.
AgentWorldBench samples trajectories from 5 frontier models across 9 real benchmarks and uses ground-truth rubric judging on 5 dimensions (1–5 each, total 25):
| Model | AgentWorldBench Score | |-------|----------------------| | Qwen-AgentWorld-397B-A17B | 58.8 | | GPT-5.4 | 58.2 | | Claude Opus 4.8 | 56.6 | | Gemini 3.1 Pro | 54.6 | | DeepSeek V4-Pro | 53.0 | | Qwen3.6-Plus | 50.8 |
Qwen-AgentWorld beats GPT-5.4 and Claude Opus 4.8 at "imagining the world"—in zero-shot mode, predicting outputs of environments it has never seen.
Two Usage Paradigms
Paradigm 1: Decoupled Environment Simulator
Use the model as a standalone environment for RL training:
Counter-intuitive but logical: like pilots training on extreme weather in simulators, injected failure modes prepare agents for rare-but-fatal scenarios.
Paradigm 2: Unified Agent Foundation Model (Warm-up)
Treat world-model training as an agent model's "preschool": first learn to *predict* the world, then learn to *act* in it.
The model learns general logic of "how environments respond to actions," not just Terminal commands. A variant even trains search agents in fully fictional but self-consistent environments, preventing memorization of real search API behavior while still generalizing to real search tasks.
In One Sentence
> Qwen-AgentWorld isn't about making models smarter—it gives models an imagination. It lets AI simulate the world in its head, not through pixels or vectors but through language. Once mature, this capability will rewrite the cost structure, safety, and scalability of agent training.