Agents That Simulate First, Act Second: Qwen-AgentWorld and the Transfer Effect of World Modeling

Agents That Simulate First, Act Second: Qwen-AgentWorld and the Transfer Effect of World Modeling

Researchers at Alibaba's Qwen team built an AI agent training environment out of completely invented facts. The fictional world contained databases of nonexistent people, companies, and events, designed so that no answer could be retrieved from the agent's parametric memory. Agents trained against this fictional environment, then evaluated on real web research tasks, scored 50.3% on the WideSearch benchmark. Agents trained against a live search engine scored 45.6%.

The paper describing these experiments, released June 23, 2026, introduces Qwen-AgentWorld, two open-weight models at different scales (35B-A3B and 397B-A17B) trained to function as language world models (LWMs): systems that predict what a computing environment will return in response to an agent's action, rather than deciding which action to take. This is a different problem than most current AI agent research addresses, and the results suggest it may be a more consequential one than it looks.

The Missing Half of the Agent Loop

Current agent research has focused almost entirely on the policy side of the agent-environment interaction: given what the agent observes, what should it do next? The world model side, given what an agent does, what does the environment return, has received considerably less attention. Qwen-AgentWorld frames this as a gap, and a paper by Richens et al. (2025) cited in the work argues theoretically that any agent capable of generalizing across a sufficiently broad range of tasks must have learned a world model. The Qwen team's experiments are an empirical test of that claim.

The model covers seven domains: MCP tool calls, web search, terminal sessions, software engineering tasks, Android UI navigation, web browser interaction, and desktop operating system control. For text-based domains, the model predicts outputs like terminal stdout, JSON API responses, and file contents. For GUI domains, it predicts the resulting screen state as HTML, which can be rendered into a screenshot.

Training proceeded through three stages on a dataset of more than 10 million environment interaction trajectories. Continual pre-training (CPT) on these trajectories, supplemented by domain-specific corpora covering law, medicine, finance, and other fields, injected general knowledge of how different systems behave. Supervised fine-tuning (SFT) activated explicit step-by-step reasoning about state transitions before each prediction. Reinforcement learning (RL) used a hybrid reward combining a five-dimensional rubric scored by an LLM judge with rule-based binary correctness signals, sharpening simulation fidelity at the level of individual predictions.

Benchmark Results

To evaluate world models, the team built AgentWorldBench, a new benchmark constructed from real environment interactions of five frontier models, including Claude Opus 4.6, Claude Opus 4.8, GPT-5.4, and Gemini 3.1 Pro, running on nine established agent benchmarks such as Terminal-Bench 1.0 and 2.0, SWE-Bench Verified, and OSWorld-Verified. The benchmark contains 2,170 turn-level samples across all seven domains, with ground-truth observations drawn from actual environment execution rather than model-generated outputs.

Qwen-AgentWorld-397B-A17B achieved an overall average of 58.71 on this benchmark, compared to 58.25 for GPT-5.4 and 56.59 for Claude Opus 4.8. The advantage was concentrated in text-based domains, particularly Terminal (57.73 vs. 53.69 for GPT-5.4) and software engineering (68.49 vs. 66.29), where accurate prediction requires tracking code execution state across many tool calls. On GUI domains, Claude Opus 4.8 and 4.6 led, a gap the authors attribute to advantages from multimodal pre-training that text-only world modeling does not replicate.

The internal baseline is instructive. Qwen3.6-Plus, from the same architecture family but without world-model training, scored 50.81 overall. Qwen-AgentWorld-397B-A17B scored 58.71. The eight-point difference is attributable to the three-stage training pipeline. All results reported here are vendor-reported by the Qwen Team; the paper is a preprint and has not undergone independent peer review.

Factuality was the lowest-scoring dimension throughout training and across all models. It also showed the largest relative improvement from RL, rising 11.3% over its initial value. The authors read this as indicating that factual world knowledge is the hardest aspect of environment simulation, and that RL pushes the model toward more accurate environment responses rather than merely cleaner formatting.

Two Ways World Modeling Improves Agents

The paper explores two distinct applications of the LWM. The first uses it as a standalone environment simulator, letting agents train via RL against simulated environments rather than real ones. The second integrates world-model training directly into the agent's own learning.

For the simulator application, the team synthesized 4,000 OpenClaw environments, an open-source personal AI agent platform entirely absent from LWM training, using a small pool of real interaction traces as seeds. Agents trained via RL in these simulated environments improved their Claw-Eval score from 65.4 to 69.7 (+4.3) and QwenClawBench from 47.9 to 55.0 (+7.1), without domain-specific adaptation.

The larger gains came from controllable simulation, where natural-language instructions shape what the simulated environment returns at each turn. These instructions can inject targeted conditions, intermittent API errors, incomplete intermediate results that require follow-up calls, responses that withhold information until the agent takes additional steps. In MCP tool-use experiments, controllable simulation lifted MCPMark scores by 12.3 points over the baseline. Without the control instructions, standard simulation produced no meaningful improvement and actually dropped performance slightly on one benchmark. Controllability was a prerequisite, not an enhancement.

The fictional-world search experiments illustrate how far this can extend. The team built 1,000 self-contained fictional environments, each anchored by a relational database of 300 to 500 internally consistent fictional facts, covering invented geographies, companies, and events. The world model's simulation instructions required search snippets to withhold complete answers, forcing agents to reformulate queries and extract full page content across multiple retrieval rounds.

Agents trained in these fictional environments increased their web_extractor calls from 2.5 to 4.0 per trajectory. Agents trained against real search engines decreased their extractor calls from 2.5 to 1.5, because real snippets often provided sufficient information and the agent learned not to bother looking further. The behavioral divergence was the mechanism behind the performance gap. The fictional environment was more adversarial by design, and that adversarial design shaped better retrieval behavior.

The Foundation Model Effect

The second application is where the results become harder to attribute to straightforward skill transfer. The team applied RL with world-model training data to a base model and then evaluated that model directly on multi-turn, tool-calling agentic tasks, with no additional fine-tuning. The LWM training data consists of single-turn, non-agentic trajectories with no tool calls. The downstream evaluations involve multi-turn tool use across domains the LWM was never trained on.\

Across seven benchmarks, the average improvement was approximately 9 points. Terminal-Bench 2.0 accuracy rose from 33.25 to 39.55. SWE-Bench Verified improved from 64.5 to 67.9. Claw-Eval, an out-of-domain benchmark absent from LWM training, gained 11.3 points. BFCL v4, a function-calling evaluation also absent from training, gained 9.0 points.

Examining reasoning traces from the model during agentic evaluation revealed a consistent pattern: after LWM RL, the model generated explicit predictions of environment responses before executing actions, using its internal model to anticipate outcomes and revise its plan before committing.

A documented case from Terminal-Bench 2.0 shows the pre-LWM model incorrectly predicting that transport routing in Postfix would precede recipient validation, causing it to attempt the wrong fixes repeatedly until timeout. The post-LWM model correctly predicted that recipient validation occurs first, targeted the correct configuration file, and solved the task. Internal prediction accuracy across analyzed turns rose from 69.9% to 78.3% after LWM training.\

The cross-domain generalization experiments add another layer. When Stage 3 RL ran on Terminal data alone, performance on three held-out text domains improved simultaneously within the first 10 steps: SWE gained 11.5 points, Search gained 11.8 points, and MCP gained 5.0 points. Terminal shell commands and MCP tool calls differ substantially in syntax, state representation, and response structure. The gains suggest the model is learning something more general than domain-specific output formats, closer to a capacity for reasoning about how actions propagate through stateful systems across turn boundaries.

\

What This Suggests for Agent Development

The fictional-world result captures something that has practical implications for organizations building on agentic AI. An agent trained entirely on invented facts, in environments with no connection to any real search index or database, developed retrieval strategies that transferred to real information-seeking tasks and outperformed agents trained on real data. The fictional environment worked because it was designed to make certain behaviors necessary. A live search engine does not require agents to issue follow-up extraction calls when snippets are sufficient, so agents trained there learn not to bother. The simulated environment made thoroughness a requirement, and that requirement shaped the agent.

The open-weight release of both model sizes makes these techniques available to organizations building on self-hosted infrastructure. The research also introduces AgentWorldBench as a public evaluation framework, adding a reference point that did not previously exist for measuring world-model quality across multiple agent domains.

The Tony Hawk Paradox framing fits the fictional-world result closely: capability that develops inside controlled, artificial environments, invented worlds with no real-world referent, transfers directly into improved performance on the real tasks that matter economically. In this case, the simulation was not a proxy for the real thing. It was a more demanding version of it.