What is a world model?

English

The simplest definition is: a world model is a learned model of environment dynamics. It looks at observations, compresses them into an internal state, and predicts what may happen next.

That definition matters because it separates world models from two things people often confuse with them. A large language model predicts text. A video model generates plausible visual sequences. A world model should be useful when an action changes the world: if the agent turns, pushes, drives, waits or changes a goal, the model should update the future accordingly.

Why this became important again

The 2018 World Models project made the idea memorable: train a compact internal model of the environment, then train an agent inside that model's imagined rollouts. Newer systems widened the term. Sora framed video generation as a path toward world simulators. DeepMind's Genie line pushed interactive generated environments. NVIDIA Cosmos treats world foundation models as infrastructure for physical AI. Meta's V-JEPA route emphasizes predictive representations rather than visible video output.

What this definition proves

• Prediction is the center of the concept: the model should estimate future state, not only produce output.
• Actions matter: a stronger world model should change the rollout when an agent intervenes.
• Usefulness is downstream: the model should help with planning, training, testing or control.

What it does not prove

• A model is not a full world model just because it generates realistic video.
• A company using the term "world model" does not mean its system has externally verified physics.
• A good demo does not prove long-horizon stability or robot/autonomous-driving safety.

What to check before believing a claim

• Does the model only generate a plausible clip, or can it condition on actions?
• Does it preserve objects, space and causal consequences across time?
• Can a downstream agent use it for training, planning, testing or control?
• Is the claim supported by a paper, technical report, product access or only a launch demo?

Common misconceptions

• "World model" is not one architecture. It can mean latent dynamics, video rollouts, 3D spaces or physical-AI infrastructure.
• Better video is not automatically better world modeling. It may improve consistency without enabling action-conditioned planning.
• Simulation is not the same as truth. A simulator is useful only when its errors are understood and bounded for the task.

Next reading

中文

最简单的定义是：世界模型是一个学习环境动态的模型。它观察世界，把观察压缩成内部状态， 然后预测接下来可能发生什么。

这个定义重要，是因为它能把世界模型和两个常见混淆对象分开。大语言模型预测文字； 视频模型生成看起来合理的视觉序列；世界模型应该在“行动改变世界”时仍然有用： 如果智能体转向、推动、驾驶、等待或改变目标，模型应该能相应更新未来状态。

为什么这个概念重新变重要

2018 年的 World Models 项目让这个想法变得清晰：先训练一个紧凑的内部环境模型， 再让智能体在这个模型想象出的轨迹里训练。后来的系统把这个词扩展了。Sora 把视频生成 放进通向世界模拟器的叙事；DeepMind Genie 系列强调可交互生成环境；NVIDIA Cosmos 把世界基础模型当作物理 AI 基础设施；Meta V-JEPA 路线则强调预测表征，而不是生成可观看视频。

这个定义能证明什么

• 预测是核心：模型应该估计未来状态，而不只是生成输出。
• 行动很重要：更强的世界模型应该在智能体干预后改变推演结果。
• 价值看下游：模型应该能帮助规划、训练、测试或控制。

这个定义不能证明什么

• 一个模型能生成逼真视频，不等于它就是完整世界模型。
• 公司使用“world model”这个词，不等于它的物理理解已经被外部验证。
• 演示好看，不等于长周期稳定，也不等于能安全用于机器人或自动驾驶。

相信一个世界模型声明前，先检查什么

• 它只是生成一段合理视频，还是能接受动作条件？
• 它能否在时间中保持物体、空间和因果后果？
• 下游智能体能否用它训练、规划、测试或控制？
• 这个声明来自论文、技术报告、可访问产品，还是只有发布演示？

常见误解

• 世界模型不是单一架构。它可能指潜空间动态、视频推演、3D 空间，或物理 AI 基础设施。
• 视频更好不等于世界模型更强。视频一致性可能提高，但仍未必支持行动条件规划。
• 模拟不等于真实。模拟器只有在误差被理解并能服务具体任务时才有价值。

下一步阅读

Evidence table / 证据表