Qwen-AgentWorld release: language world model for seven domains and its impact on general agents

TL;DR

Qwen releases Qwen‑AgentWorld, a native language world model that simulates seven text‑ and GUI‑based agent environments within a single model and achieves state‑of‑the‑art simulation quality on the new AgentWorldBench benchmark. The model enables two complementary ways to improve general agents: as a controllable simulator for reinforcement learning, and as a unified foundation model that internalizes next‑state prediction.

Overview of Qwen‑AgentWorld

What it is – Qwen‑AgentWorld is a language world model (LWM) trained to predict the next observation of an environment given the current interaction history and an agent’s action. Unlike prior approaches that adapt a general‑purpose LLM after the fact, Qwen‑AgentWorld incorporates environment modeling as an explicit objective from the start of continual pre‑training (CPT) through supervised fine‑tuning (SFT) and reinforcement learning (RL).

Scope – The model covers seven interaction domains:

  • Text‑based: Terminal, Search, MCP (tool‑call API), Software‑Engineering (IDE/code editing).
  • GUI‑based: Web browser, Desktop OS, Android UI. For GUI domains the model predicts structured renderable code (HTML, accessibility‑tree XML, UI hierarchy) instead of raw pixels, allowing pure‑text simulation.

Benchmarks – Alongside the model the team releases AgentWorldBench, a seven‑domain evaluation suite that pairs each simulated trajectory with a ground‑truth observation from a real environment. Scoring uses a five‑dimensional rubric (format, factuality, consistency, realism, quality) judged by an LLM.

Training Pipeline

1. Continual Pre‑Training (CPT)

  • Injects environment knowledge from >10 M real interaction trajectories collected from sandboxes, emulators, and open‑source benchmarks.
  • Augments trajectories with specialized world‑knowledge corpora (industrial control, cybersecurity, law, medicine, finance, current affairs).
  • Applies turn‑level information‑theoretic loss masking to focus learning on turns that contain genuine environment information.

2. Supervised Fine‑Tuning (SFT)

  • Introduces an explicit next‑state prediction pattern using `` blocks.
  • Selects high‑quality thinking trajectories via rejection sampling, yielding 7,094 SFT samples.

3. Reinforcement Learning (RL)

  • Sharpens fidelity with GSPO RL.
  • Rewards combine a rubric‑based LLM judge (multi‑dimensional quality) and rule‑based verifiers for domains where exact correctness is programmatically checkable.

Performance on AgentWorldBench

Model Overall Score Notable Domains
Qwen‑AgentWorld‑397B‑A17B 58.71 (highest) Terminal, SWE (large gains)
GPT‑5.4 58.25 –
Claude Opus 4.8 – –
Gemini 3.1 Pro – –

At the 35 B‑A3B scale, the three‑stage pipeline raises the overall average from 47.73 to 56.39 (+8.66 points), surpassing Claude Sonnet 4.6 (56.04). Gains are consistent across both text and GUI domains.

Emergent Reasoning Patterns

Analysis of 129 thinking traces across four text domains revealed three distinct patterns:

  1. Deliberative self‑correction – The model inserts “Wait!” interrupts to revise intermediate predictions, averaging 10.4 interrupts per turn.
  2. Information leakage prevention – In Search, the model withholds target answer snippets when the query is unrelated, mimicking a theory‑of‑mind behavior.
  3. Multi‑step causal reasoning – Predicting complex pipelines (e.g., curl -s localhost:3000 | python3 -m json.tool) requires chaining six logical steps, which the model successfully performs.

Paradigm I: Decoupled Simulation for RL

Why simulate?

  • Scalability & controllability: LWMs generate unlimited turn‑level interactions without needing sandbox infrastructure, and can inject rare or adversarial perturbations.
  • Internal planning: Agents can use a world model to mentally simulate outcomes before acting, a capability absent from pure policy learning.

Key Findings

  • Zero‑shot generalization: Simulating 4 k OpenClaw environments (unseen during training) yields +4.3 on Claw‑Eval and +7.1 on QwenClawBench.
  • Controlled simulation matters: Uncontrolled Sim‑RL provides negligible gains; adding natural‑language control instructions improves MCPMark by +12.3 and Tool Decathlon by +3.7.
  • Surpassing real‑environment RL: On WideSearch, controllable Sim‑RL reaches 50.3 % F1 (step 60) versus 45.6 % for RL trained on a live search engine.
  • Behavioral shaping: Controllable simulation drives agents to increase web_extractor calls, reflecting learned reliance on deeper snippet extraction.

Fictional‑World Experiments

Creating 1 k self‑contained fictional databases forces agents to rely on the simulated search tool. Controlled Sim‑RL improves Item‑F1 from 34.02 to 50.31 (+16.29) and Row‑F1 from 13.72 to 24.21 (+10.49) for the 35 B model.

Paradigm II: Unified Agent Foundation Model

In this paradigm the same LWM both selects actions and predicts next observations, embedding world‑modeling as a meta‑reasoning skill.

Transfer Results

Training the LWM on a single‑turn, non‑agentic RL task (no tool calls) yields cross‑task generalization to multi‑turn, tool‑calling benchmarks across all seven domains, including three out‑of‑distribution sets:

Benchmark (out‑of‑domain) Base Score LWM‑RL Score Δ
Claw‑Eval 33.3 39.6 +6.3
QwenClawBench 64.5 67.9 +3.4
BFCL v4 42.2 47.4 +5.2
WideSearch (F1 Item) 33.4 46.2 +12.8
… … … …
These improvements arise without any RL fine‑tuning on the target agentic tasks, demonstrating that next‑state prediction can serve as a powerful foundation skill.

Deployment Details

  • Model sizes: 35 B‑A3B (MoE, 3 B active parameters) and 397 B‑A17B.
  • Context window: 256 K tokens.
  • Serving: Compatible with SGLang and vLLM. Example commands:
# SGLang
python -m sglang.launch_server \
  --model-path Qwen/Qwen-AgentWorld-35B-A3B \
  --port 8000 \
  --tensor-parallel-size 4 \
  --context-length 262144 \
  --reasoning-parser qwen3

# vLLM
vllm serve Qwen/Qwen-AgentWorld-35B-A3B \
  --port 8000 \
  --tensor-parallel-size 4 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --trust-remote-code
  • Evaluation pipeline: eval/eval.py runs inference, LLM‑based judging, and aggregation across the five rubric dimensions.

Implications for General Agents

  1. Complementary scaling axis – LWMs provide a controllable, scalable simulation layer that augments, rather than replaces, real‑environment interaction.
  2. Unified reasoning – Embedding next‑state prediction within the agent creates a mental‑simulation capability akin to reflective planning, strengthening instruction following, long‑context handling, and domain knowledge utilization.
  3. Cross‑domain transfer – Training on diverse trajectories yields shared modeling skills that benefit unseen domains, reducing the need for domain‑specific data.
  4. Research roadmap – Qwen‑AgentWorld demonstrates that language‑only world modeling can reach and exceed frontier model performance, opening a path toward more capable, general‑purpose autonomous agents.

Citation

@article{zuo2026qwen,
  title   = {Qwen‑agentworld: language world models for general agents},
  author  = {Zuo, Yuxin and Xiao, Zikai and Sheng, Li and Huang, Fei and Tu, Jianhong and Liu, Yuxuan and Tang, Tianyi and Hu, Xiaomeng and Su, Yang and Lan, Qingfeng and others},
  journal = {arXiv preprint arXiv:2606.24597},
  year    = {2026}
}

Sources