Imagine playing an open-world game. How does a traditional LLM agent handle it? Every single step requires actually connecting to the server, opening the page, executing the command, and waiting for the result — like a person who, every time they wonder "what happens if I jump from here," must actually jump.
What Qwen-AgentWorld does is build a simulator inside the AI's head. When the agent asks "what happens if I grep this file?", the model doesn't need to actually execute it — it can "imagine" the terminal output in its head. When the agent asks "what happens if I click this button?", the model can predict the next screen's UI state.
That is the essence of a Language World Model (LWM) — not making the AI "do" actions, but letting the AI "rehearse" their outcomes.
---
Why It Matters: AI's "Imagination"
Humans have a core capability when making decisions: simulating consequences in the mind before acting. In chess, we think "if I make this move, how will my opponent respond?"; when driving, "if I change lanes now, how will the car on the right react?"
Traditional AI agents lack this ability. Their loop is: observe → think → act → wait for real feedback → think again. Every step requires interaction with the real environment — slow, expensive, irreversible. If the environment is a surgical robot or a trading system, the cost of trial and error becomes unbearable.
The breakthrough of Qwen-AgentWorld: it turns the language model itself into an interactive virtual environment. You can train agents inside this "environment," letting them try various strategies and see which works — and this "environment" runs entirely inside the model, no real servers, web pages, or phones needed.
---
Two Models: Light Cavalry and Heavy Armor
Alibaba released two variants:
- 35B-A3B: 35B total parameters, 3B activated (MoE). For ordinary users — runs on a single A100, deployable by SMBs.
- 397B-A17B: 397B total parameters, 17B activated (MoE). For heavy workloads — peak performance, requires multi-GPU interconnect.
- Terminal: Linux command line, predicting cat/grep/pip output
- SWE: software engineering, predicting tracebacks or diffs after code execution
- Search: search engines, predicting result pages
- MCP: tool calls, predicting API-returned JSON
- Web: web browsers, predicting DOM changes after clicks
- Android: phone UIs, predicting UI hierarchy changes after touches
- Desktop OS: desktop systems, predicting window states after mouse/keyboard input
- 9 points from a rubric judge: 5 dimensions (format, factuality, consistency, plausibility, quality), each scored 1–5 by an LLM judge. Rich multi-dimensional feedback, but vulnerable to reward hacking (the model learning to flatter the judge).
- 1 point from a rule verifier: executable verifiers, 0/1 correct-or-not. The hard anchor preventing the model from scoring points with "silver-tongued" answers.
- Sampling trajectories from 5 frontier models (Claude Opus 4.6, etc.) across 9 real benchmarks
- Evaluating with ground-truth rubric judging across 5 dimensions: format, factuality, consistency, plausibility, quality
- Each dimension scored 1–5, total out of 25 (scaled per trajectory)
- Scalable: simulate thousands of environments in parallel, no real servers needed
- Controllable: make the simulator "deliberately" return errors, pagination, partial failures — edge cases rarely encountered in real environments but essential to master
- Outperforms real-environment training: on OpenClaw, Sim RL beats Real RL by +4.3 (Claw-Eval) to +7.1 (QwenClawBench); in controlled simulations, MCPMark improves +12.3 and WideSearch +16.3
- Across-the-board gains on 7 agent benchmarks
- RL trained only on Terminal data yields SWE gains of +11.5, Search +11.8, MCP +5.0 — cross-domain emergence
The dual-spec release suggests Qwen is betting that "world models" will move from the lab to the production line. Not every company needs the 397B behemoth, but the 35B light cavalry lets far more people participate.
---
Three-Stage Forging: From "Having Seen" to "Able to Imagine"
Training a world model isn't just "feed it lots of data." Qwen-AgentWorld uses a three-stage pipeline, each step with a clear goal:
Stage 1: CPT (Continual Pretraining) — Let the Model "See the World"
Train on 10M+ real environment interaction trajectories covering 7 domains:
These trajectories weren't written by humans — they were produced by real agents running in real environments. Through CPT, the model "sees" countless action–outcome pairs and builds basic intuitions about how the world works.
A clever engineering detail: information-theoretic loss masking. In tool-call trajectories, many turns are "fluff" — e.g., an echo command 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 "fluff" turns. This preserves historical context (later turns depend on it) without wasting compute teaching the model to parrot.
Stage 2: SFT — Activating "Predict the Next State" Thinking
In CPT the model implicitly learns what follows an action; SFT explicitly teaches it "you are predicting the next state." The model starts producing long chain-of-thought reasoning: "what is the current state → what does the user action mean → based on physics/system logic, what should the result be → output the prediction."
SFT used rejection sampling: 3 rollouts per query, scored by a judge model, keeping only the highest-scoring rollouts above threshold. From 10,250 candidate queries, only 7,094 survived (69.2% retention). Good steel on the blade's edge — for SFT, data quality matters more than quantity.
Stage 3: RL — Sharpening Simulation Fidelity
After SFT the model "can imagine," but imprecisely. RL refines it with a hybrid reward:
The 9:1 ratio is deliberate — rubric provides the signal, rule provides the floor. The paper also flags a pitfall: expanding one trajectory into many turn-level samples during RL creates long shared prefixes, causing reward collapse (samples too similar, nothing learned). The fix: expand only one turn per trajectory, ensuring each training sample has an independent prediction target.
---
AgentWorldBench: What Makes You a World Model?
Evaluating world models is hard. Traditional metrics (BLEU, ROUGE) only measure surface similarity, but a world model's output may be "same meaning, different wording," or contain IDs and timestamps irreproducible in the real environment.
The Qwen team built AgentWorldBench:
Results:
| 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 |
On "imagining the world," Qwen-AgentWorld beats GPT-5.4 and Claude Opus 4.8. Note this is zero-shot — no environment-specific fine-tuning; the model directly predicts outputs of environments it has never seen.
---
Two Usage Paradigms: Simulator and Coach
The paper's most strategically interesting contribution is two complementary paradigms:
Paradigm 1: Decoupled Environment Simulator
Use Qwen-AgentWorld as a standalone environment in which agents train via RL. Advantages:
This defies intuition: "how can simulated beat real?" The answer — the simulator can generate challenge scenarios rare in real environments. Just as pilots practice extreme weather in simulators, which is safer than only flying in clear skies. The model can "inject API errors," "force pagination," or "return partial failures," teaching agents to handle situations that are uncommon in reality but fatal when they occur.
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 the world." Effects:
The paper also mentions an intriguing variant: fictional-world training. Training a search agent in a fully fictional but self-consistent environment prevents it from memorizing real search-engine parameters to cheat. The trained agent still generalizes to real search tasks — because it learned "search logic," not "Google's specific API behavior."
---
In One Sentence
> Qwen-AgentWorld isn't making the model smarter — it's giving the model an imagination. It gives AI, for the first time, the ability to 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.
---
Paper: https://arxiv.org/abs/2606.24597 Code and models: https://github.com/QwenLM/Qwen-AgentWorld