A world model predicts environment dynamics based on current observations and actions, serving as a core cognitive mechanism for reasoning and planning. In this work, we investigate how world modeling based on language models can further push the boundaries of...

Qwen-AgentWorld: Language World Models for General Agents

Episode 0044 — DTF:FTL | Daily Tech Feed: From The Labs

What This Paper Does

Qwen-AgentWorld, from Alibaba's Qwen team, builds the missing half of the AI agent equation: a language world model — a system that predicts what happens next in an environment when an agent takes an action.

Current AI agent research has focused almost entirely on the policy side: what action should the agent take? Qwen-AgentWorld addresses the complementary question: given the current state and an action, what is the next state? This is the world model. The paper argues, backed by a 2025 theoretical proof (Richens et al.), that any agent capable of generalizing across a broad range of tasks must have learned a world model.

The result is two open-weight models — Qwen-AgentWorld-35B-A3B (released; 35B parameters, 3B active, Mixture-of-Experts) and Qwen-AgentWorld-397B-A17B (benchmark-evaluated) — capable of simulating seven categories of agent environments through long chain-of-thought reasoning.

The Seven Domains

The model simulates all of the following within a single unified framework:

MCP (Model Context Protocol tool calls)
Search (web search and extraction)
Terminal (shell commands, bash)
SWE (software engineering: read/edit/bash workflows)
Android (touch/swipe/type on UI view hierarchies)
Web (click/navigate via accessibility trees)
OS (mouse/keyboard on desktop environments)

For the three GUI domains, observations are represented as textual accessibility trees and UI view hierarchies rather than pixel frames — making them tractable for language model training.

How It Was Trained

Three-stage pipeline — "CPT injects, SFT activates, RL sharpens":

Continual Pre-Training (CPT): Trained on 10M+ real-world interaction trajectories collected from three sources: a dedicated agent infrastructure running automated tasks across all seven domains, open-source interaction traces (terminal recordings, agentic tool-call logs), and in-house Alibaba agentic trajectories. CPT injects environment dynamics without chain-of-thought reasoning.
Supervised Fine-Tuning (SFT): Activates next-state prediction as an explicit thinking pattern — the model learns to reason through what the environment will return before generating its prediction.
Reinforcement Learning (RL): Sharpens fidelity with a hybrid reward system combining rubric-based scoring (open-ended quality dimensions) and rule-based verifiers (deterministic checks).

Data pools across the three stages are strictly disjoint. The RL pool alone contains 92,308 trajectories averaging 13.4 turns each.

AgentWorldBench

A new evaluation benchmark built from real environment interactions of five frontier models on nine established agent benchmarks, including Terminal-Bench 1.0 and 2.0, OSWorld-Verified, and others. Evaluation uses rubric judging across five dimensions. All eval trajectories are out-of-distribution for the trained models.

AgentWorldBench results (overall score, higher is better):

Model	Overall
Qwen-AgentWorld-397B-A17B	58.71
GPT-5.4	58.25
Claude Opus 4.6	57.80
Claude Opus 4.8	56.59
Claude Sonnet 4.6	56.04
Qwen-AgentWorld-35B-A3B	56.39
Qwen3.5-35B-A3B (no LWM)	47.73

The 35B model with LWM training shows a +8.66 point improvement over the same model without it.

Two Ways to Use a World Model

Paradigm 1: Decoupled Environment Simulator

Use the world model to simulate environments for agentic RL training, eliminating the need for real-environment access. Key results:

Generalizable simulation: Sim RL on 4,000 out-of-distribution OpenClaw environments yielded +4.3 on Claw-Eval and +7.1 on QwenClawBench vs. real-environment RL with a weaker simulator.
Controllable perturbations (MCP): Injecting targeted adversarial conditions (e.g., hidden answers, degraded tool responses) during training: +3.7 on Tool Decathlon, +12.3 on MCPMark.
Fictional-world construction (Search): Agents trained entirely in invented, self-consistent fictional search worlds: +16.29 on WideSearch F1 Item, +10.49 on WideSearch F1 Row — surpassing real-environment training.

The fictional-world result is particularly striking. Self-consistency of the simulated world, not factual accuracy, is what matters for generalization.

Paradigm 2: Unified Agent Foundation Model

Use LWM training as a warm-up or auxiliary training stage before downstream agentic RL. The world model acquaints the agent with environment dynamics before it has to act.

Agent performance gains (35B model, LWM RL warm-up vs. SFT baseline):

Benchmark	Baseline	w/ LWM RL	Gain
Terminal-Bench 2.0	33.25	39.55	+6.30
SWE-Bench Verified	64.47	67.86	+3.39
SWE-Bench Pro	42.18	47.42	+5.24
WideSearch F1 Item	33.38	46.17	+12.79
Claw-Eval	53.60	64.88	+11.28
QwenClawBench	39.76	49.43	+9.67
BFCL v4	62.29	71.25	+8.96

Gains appear across in-domain and out-of-domain benchmarks. Three of the seven benchmarks are entirely outside the LWM training distribution.

Why This Matters

The open-weights angle: Qwen is an Alibaba project. The 35B-A3B model weights and AgentWorldBench dataset are publicly released on HuggingFace. A Chinese industrial lab releasing competitive open-weight models continues to compress the gap between proprietary frontier systems and what any researcher or developer can run.

The simulation unlock: If you can simulate environments accurately enough to train real agents, you can scale RL training without scaling real-world compute infrastructure. Every shell command, every API call, every GUI tap becomes synthetically reproducible. The fictional-world result suggests the bar for "accurate enough" may be lower than expected — internal consistency matters more than ground truth.

The missing piece argument: The theoretical backing (Richens et al. 2025: generalization requires world models) reframes this as a necessary research direction, not a nice-to-have. If that proof holds, world models are not optional.

The open questions: Does this transfer to pixel-based environments? How does simulation fidelity degrade for rare or adversarial states? The 397B model is not publicly released — the benchmark-beating number comes from the closed model.

Listen

44: Qwen-AgentWorld: Language World Models for General Agents

Show Notes