Can language models learn to simulate agent environments?

The agent-environment loop has two components: the policy (states → actions) and the world model ((states, actions) → next states). Qwen-AgentWorld's framing is that LLM-agent research has obsessed over the policy and almost entirely neglected the world model — and that this is a load-bearing gap, citing Richens et al.'s result that any agent generalizing broadly must have learned a world model. The contribution is a native language world model (35B-A3B and 397B-A17B) that simulates agentic environments across seven domains via long chain-of-thought, trained on 10M+ real interaction trajectories through a three-stage "CPT injects, SFT activates, RL sharpens" recipe targeting next-state prediction.

Two claims earn their keep. First, as a decoupled simulator, controllable simulation beats both uncontrolled simulation and — strikingly — real-environment training on three agentic benchmarks, suggesting a trained world model can manufacture cleaner, more targeted experience than reality supplies. Second, as a foundation, LWM warm-up improves downstream agent performance across all seven tasks via cross-domain transfer, positioning next-state prediction as a transferable pretraining objective for agents the way next-token prediction is for language.

This refines the conceptual debates. It operationalizes What should a world model actually be designed to do? — though notably Qwen-AgentWorld's mechanism is next-state prediction, so it either contradicts that essay's "not next observation" claim or shows that purposeful simulation can be built out of next-state prediction at sufficient scale and reasoning depth. It also bears on Do LLMs actually have world models or just facts?: a model that controllably simulates state transitions is reaching for the mechanistic (Sense 2) reading, but benchmark success on seven curated domains does not prove a compact generative model of mechanisms versus sophisticated interpolation over 10M trajectories. The strongest counterargument is the simulation-fidelity ceiling: an agent trained on simulated rather than real experience inherits the world model's errors, and "surpassing real-environment training" may hold only where the simulator's blind spots happen not to matter.