Qwen-AgentWorld: Language World Models for General Agents

TL;DR

Qwen-AgentWorld is a pair of language world models (35B-A3B and 397B-A17B MoE) trained to simulate the dynamics of agentic environments — predicting the next-state of browsers, code interpreters, terminals, etc. given a current observation and an action — via long chain-of-thought reasoning. Trained on 10M+ trajectories across 7 domains through a three-stage pipeline (CPT injects general world-modeling capability, SFT activates next-state-prediction reasoning, RL with hybrid rubric+rule rewards sharpens simulation fidelity), the models outperform existing frontier models on AgentWorldBench, the paper’s own benchmark constructed from 5 frontier models on 9 established agentic benchmarks. Beyond serving as foundation models for environment simulation, the paper shows the same recipe acts as (i) a decoupled simulator that supports scalable agentic RL with gains exceeding real-environment training and (ii) a warm-up that improves downstream agent performance across 7 agentic benchmarks when used as a unified agent foundation model.

Key claims

Qwen-AgentWorld introduces language world models — LLMs trained to predict next-state observations in agentic environments via long chain-of-thought reasoning — covering 7 domains [Abstract].
Two model sizes are released as MoE architectures: Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B [Abstract].
Training uses 10M+ environment interaction trajectories collected from real-world environments across the 7 domains [Abstract].
The three-stage training pipeline is functionally specialized: CPT injects general-purpose world modeling capabilities from state-transition dynamics and augmented professional corpora; SFT activates next-state-prediction reasoning; RL with a tailored framework and hybrid rubric-and-rule rewards sharpens simulation fidelity [Abstract].
AgentWorldBench is constructed from real-world interactions of 5 frontier models on 9 established benchmarks, providing a comprehensive evaluation suite for language world models [Abstract].
Qwen-AgentWorld significantly outperforms existing frontier models on AgentWorldBench [Abstract].
Used as a decoupled environment simulator, Qwen-AgentWorld supports scalable and controllable simulation of thousands of real-world environments for agentic RL, yielding gains that exceed training in real environments alone [Abstract].
Used as a unified agent foundation model, world-model training acts as an effective warm-up that improves downstream performance across 7 agentic benchmarks [Abstract].

Method

The paper has two layers. The foundation-model layer trains LLMs (35B-A3B and 397B-A17B MoE backbones, presumably from Qwen3 base) to simulate environment dynamics: given a current observation and an action, predict the next observation. Critically, the prediction goes through long chain-of-thought reasoning — the model thinks about what should happen before emitting the next state — which differentiates the recipe from naïve trajectory imitation and aligns with the Qwen-series thinking-model conventions. The training pipeline is three-stage: continued pretraining (CPT) on 10M+ real environment-interaction trajectories spanning 7 agentic domains plus augmented professional corpora to inject general-purpose world-modeling priors; supervised fine-tuning (SFT) to activate the next-state-prediction reasoning chain; reinforcement learning with a custom framework and a hybrid rubric-and-rule reward — rubrics (presumably LLM-judged criteria like state correctness, format, semantic plausibility) plus rules (exact-match / structural checks against ground truth where they apply) — to sharpen simulation fidelity.

The system layer demonstrates two use cases. As a decoupled environment simulator, Qwen-AgentWorld replaces or augments real environments during agent RL training: thousands of simulated environments run in parallel against an agent policy, and the gain reportedly exceeds training in real environments alone (the simulator’s controllability — reproducibility, scaling, and selection — beats real-environment fidelity at the margins where RL is most data-hungry). As a unified agent foundation model, the world-model training itself acts as a pretraining task whose by-product is a stronger general agent: warm-starting from the world-modeling checkpoint and then SFT/RL’ing on agentic tasks improves downstream performance across 7 benchmarks. AgentWorldBench, the evaluation harness, is constructed by recording real-world interactions of 5 frontier models against 9 established benchmarks and scoring whether candidate world models reproduce those transitions.

Results

Significantly outperforms existing frontier models on AgentWorldBench [Abstract]. No specific deltas disclosed at filing time (abstract-only fetch).
As a decoupled simulator for agentic RL, gains exceed real-environment training alone [Abstract] — no specific benchmark deltas disclosed.
As a unified agent foundation model, improves downstream performance across 7 agentic benchmarks [Abstract] — no specific deltas disclosed.
Code release: project repository linked from the abstract; specific URL not retrievable at filing time.

Why it’s interesting

This is the strongest filed datapoint for language world models — extending the World Foundation Models cluster from its existing pixel/video-rollout pole (Cosmos 3, Genie 3, Qwen-RobotWorld) and latent-predictive JEPA pole (V-JEPA 2.1, LeWorldModel) into a third pole: agentic environment simulation in token space. Where Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation (sibling Qwen-team release one week earlier) predicts pixels of future robot trajectories given language actions, Qwen-AgentWorld predicts the textual observations a browser/terminal/code-interpreter would return given a tool call. This makes it the natural language-side counterpart to Cosmos 3’s “world model that emits both video and action” — Qwen-AgentWorld is the model that emits the environment response.

It also lands directly in the RL Environment Platforms debate. The existing entries on that page (SETA, Toolathlon-GYM, OpenReward, AWM) all package executable environments as Docker stacks or managed services because they accept the constraint that an environment is what an environment returns under execution. Qwen-AgentWorld inverts this: a sufficiently capable LLM can simulate the environment, and the result is more scalable and controllable than real-environment rollouts at training time. This is the strongest counter-bet to the verifier-cheap / action-expensive design assumption that drives the env-as-package pattern — if the simulator is good enough, the action becomes cheap too. The closest comparison is Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning (AWM), which synthesizes 1,000 executable SQL-backed environments with verifiers; Qwen-AgentWorld instead learns a single neural simulator that covers many environments, trading explicit grounding for breadth and reasoning. The two recipes are likely complementary: AWM-style synthesized environments are an obvious data source for Qwen-AgentWorld-style world-model training.

The “world-model training as warm-up improves downstream agent performance” finding is the third interesting hook — it connects Reasoning RL and Tool-Use Agents by suggesting that learning to predict environment dynamics is a useful pretraining task for learning to act in those environments, a textual analog of the LingBot-VA “world-model-as-policy-internal-substrate” pattern from the embodied side.