World Models

World models are AI systems that learn compressed internal representations of an environment — its physics, appearance, dynamics, and rules — enabling prediction of future states and generation of novel scenarios. They represent a fundamental shift from hand-authored virtual worlds to worlds that emerge from learned understanding.

The concept originates in cognitive science: humans navigate the world using internal models that predict what happens next. If you throw a ball, your brain simulates its trajectory before it lands. World models in AI attempt the same: learn the "rules" of an environment from observation, then use that understanding to predict, plan, or generate.

Early world models were relatively simple — David Ha and Jürgen Schmidhuber's 2018 "World Models" paper trained a variational autoencoder and recurrent network to learn a latent representation of simple game environments. The agent could then "dream" — planning actions in the learned model rather than the real environment.

The field has since scaled dramatically. Google DeepMind's Genie (2024) and Genie 2 learned playable 2D and 3D environments from video, generating interactive worlds from single images. NVIDIA's Cosmos world foundation models are trained on massive video datasets to understand physical dynamics. These systems don't just predict the next frame — they model object permanence, gravity, collisions, and spatial relationships.

For game development, world models offer a provocative possibility: generating entire game levels or environments from learned priors rather than manual design. A world model trained on thousands of hours of gameplay footage could generate novel but plausible game scenarios, terrain layouts, or physics behaviors. This connects to procedural generation but operates at a fundamentally different level — generation from understanding rather than from rules.

World models are also critical for robotics and autonomous systems. A robot with an accurate world model can simulate the consequences of actions before executing them, dramatically reducing the need for real-world trial and error. Tesla's approach to self-driving, NVIDIA's Isaac platform for robotics, and various embodied AI projects all rely on learned world models for planning and simulation.

The convergence of world models with generative video, 3D scene representations, and language models points toward a future where AI systems don't just generate static content but interactive, physically coherent environments that respond to actions in real time.

2026 Developments

In 2026, world models moved from research demos toward products, and the money and attention around them shifted noticeably toward training robots.

World Labs Atlas and the AMD acquisition

On September 1, 2026, World Labs released a preview of Atlas, which it calls an "omni" world model: a multimodal autoregressive diffusion transformer that generates up to one minute of 1440p video and can also output point clouds and Gaussian splats, so its worlds can be used as 3D assets and not only watched. World Labs pitches Atlas for real-to-sim robot training, turning captures of real spaces into simulations robots can learn in. Four weeks later, on September 28, AMD announced it would acquire World Labs. Co-founder Fei-Fei Li becomes AMD's EVP and Chief Scientist. Terms were not disclosed, and the deal is expected to close by the end of 2026. It gives AMD a frontier model team to compete with NVIDIA's Cosmos and Isaac stack.

Runway GWM Worlds 2

On September 3, 2026, Runway released GWM Worlds 2, which generates real-time 720p video at 24 frames per second with 48kHz audio, steered by text prompts and camera control. Runway is candid about the trade-offs: real-time generation "trades fidelity for speed," and "long-term memory is also imperfect," meaning places can change when you look away and come back. Those remain the core unsolved problems for interactive world models.

Genie with Street View

At Google I/O on May 19, 2026, Google DeepMind expanded Project Genie so it can generate explorable worlds starting from real Street View locations, rolling out to Google AI Ultra subscribers. Grounding generation in real places is a step toward world models as a consumer product and as a source of realistic training environments.

Decart

Investor appetite for real-time generation is strong. Decart raised $300 million at a $4 billion valuation on May 18, 2026, in a round led by Radical Ventures with NVIDIA participating.

Gemini Robotics 2 and the shift to robotics simulation

In late July 2026, Google DeepMind reportedly introduced Gemini Robotics 2, aimed at whole-body control of humanoid robots. Reported success rates on complex multi-finger manipulation tasks were in the 32–44% range, a reminder of how far dexterous robotics still has to go. OpenAI, meanwhile, shut down the Sora app and API but says Sora continues as world-model research.

Taken together, these moves show where the field is heading. The first wave of world models was judged by how good the video looked. The current wave is increasingly judged by whether a robot trained inside the generated world can work in the real one. World Labs' real-to-sim pitch, AMD's purchase, NVIDIA's GR00T reference humanoid and DeepMind's robotics models all point to world models for robotics and physical AI as the main commercial use, with games and entertainment as a second market.

Language World Models as Environment Simulators

A separate line of work applies the idea to text instead of pixels. A language world model is a language model trained or prompted to predict what an environment will return after an agent acts: the next web page, a tool's output, a terminal's response. The appeal is practical. Real RL environments for agents are slow and costly to build and to reset, and some actions in them cannot be undone, whereas a learned simulator can be run in parallel, varied on demand and queried before committing to an action.

An early example used the simulator for planning. WebDreamer (Gu et al., November 2024) had an LLM act as both world model and value function for a web agent, imagining the outcome of candidate actions before taking one; the authors reported it beat reactive baselines and was 4–5 times more efficient than tree search in sandbox environments.

The larger 2026 result uses the simulator for training. Qwen-AgentWorld (Zuo et al., Qwen team, June 2026) introduces two models, Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, which the authors describe as the first language world models able to simulate agentic environments across 7 domains through long chain-of-thought reasoning. They were built from more than 10 million real environment-interaction trajectories in three stages: continued pre-training, supervised fine-tuning for next-state prediction, and reinforcement learning with hybrid rubric-and-rule rewards to sharpen fidelity. The paper reports two uses. As a decoupled simulator, the model supports agentic RL across thousands of simulated environments, with gains the authors say exceed training in real environments alone. As a warm-up stage, world-model training improved downstream results on 7 agentic benchmarks.

These are the authors' own results, evaluated in part on AgentWorldBench, a benchmark introduced in the same paper, and independent replication was not found as of October 2026. The known risk with any learned simulator applies here too: a policy trained inside it can learn to exploit the simulator's errors instead of the task, a relative of reward hacking. For that reason simulated rollouts are usually paired with checks against a real or executable environment before a result is trusted.