World Models 2026: From Cosmos, Genie to JEPA — The Divergence of Routes

The second half of 2026 has seen a noticeable acceleration in the world models space.

Over the past few months, I’ve written a lot of foundational content on this blog — from the mathematics of RSSM to hands-on DreamerV3 training, from the VLA vs world models comparison to Sim-to-Real engineering pitfalls. A reader recently asked: “If I want to keep up with the latest in world models, what should I be reading?”

In this article, I’ll walk through the most noteworthy papers and projects as of late August 2026. “Noteworthy” here doesn’t mean all of these projects launched in the second half of the year — some (like Cosmos and Genie) were first released in 2025 but continued to develop and exert significant influence in 2026. What matters more to me is: as of now, they still represent technology paths worth tracking over the next six months.

This isn’t an exhaustive survey list — it’s a curated selection from the perspective of an engineer working in this field, filtered by “how useful is this for my actual work?”

But before diving into specific projects, there’s something important that needs to be clarified first.

1. Let’s Distinguish Between Types of “World Models”

The term “world model” is being used too loosely now — video generation models call themselves world models, game engines call themselves world models, even simple prediction models claim the title. But the projects discussed below, while all called “world models,” are not the same thing.

Drawing on the definitions discussed in A Definition and Roadmap for World Models (arxiv 2607.06401), recent surveys, and my own engineering experience, I find it most useful to think of current world models as four technology paths. They are not strictly mutually exclusive categories, but rather a way to quickly locate the technical focus of different work:

A. Latent Dynamics World Model

Examples: Dreamer / RSSM

Core logic: state → action → next state, learning environment dynamics in latent space.

Goal: supporting planning, RL, control. This is the type I’ve written about most on this blog.

B. Generative Video World Model

Examples: NVIDIA Cosmos, etc.

Core logic: condition + history → future observations, generating future video frames.

Goals lean toward: data generation, simulation, prediction, perception.

C. Interactive World Model

Examples: Google Genie series

Core logic: state + action → interactive future, generating interactive futures based on actions.

Key capabilities: action-conditioned generation, temporal consistency, controllability.

D. Spatial / 3D World Model

Examples: World Labs Marble, etc.

Focus: persistent scene, geometry, spatial consistency, navigability, 3D representation.

With this framework in place, we won’t conflate different things when discussing specific projects below.

It’s worth noting that these paths are not concepts on the same axis. “Generative Video” describes the model’s generative paradigm, “Interactive” describes whether it supports action-conditioned interaction, while World Labs’s 2026 taxonomy of “Renderer / Simulator / Planner” describes the functional role of world models within a system. Different classification systems (such as the Functionality / Temporal Modeling / Spatial Representation three-axis framework from the 2025 Embodied AI survey) approach from different angles, each with merit. Real systems often fall on multiple dimensions simultaneously — understanding this is more valuable than debating “who is the real world model.”

The biggest change in “world models” in 2026 is not the emergence of a unified World Model, but rather the divergence of different world-model paradigms — each heading toward different application scenarios and evaluation criteria.

Let’s look at each category in turn.

2. Physical AI World Foundation Model Platform: NVIDIA Cosmos

Not a Model, but a Platform

NVIDIA first unveiled Cosmos at CES 2025. More precisely, Cosmos is a world foundation model platform for Physical AI, not a single video prediction model. By August 2025, NVIDIA had announced that Cosmos World Foundation Models had been downloaded over 2 million times.

It covers video generation, world state understanding, data processing, and synthetic data generation. Reducing it to “video generation” underestimates its technical ambition.

Why does this matter? I analyzed this in detail in my earlier article on synthetic data — real robot data collection is expensive and narrow in coverage, which is the core bottleneck for world model deployment. Cosmos’s approach is: use physics-aware generative models to produce large-scale synthetic training data for autonomous driving and robotics.

For practitioners: If you’re working on robotics or autonomous driving, Cosmos’s open-source models and toolchain deserve a serious look. If its models, data tools, and deployment ecosystem continue to mature, it has the potential to become important open-source infrastructure for Physical AI.

3. Interactive World Model: Google DeepMind Genie 3

Real-Time Generation ≠ Controllability

The Genie series represents Google DeepMind’s important investment in world models. Genie 3, released in August 2025, is positioned as “the first real-time interactive general-purpose world model.”

What can it do? Given an image or text description, it generates an interactive world simulation at 24fps — continuously evolving visually coherent scenes based on user action inputs. This is not pre-rendered video, nor is it a traditional exportable 3D world with complete geometry and physics engines. It’s closer to: generating an interactive, action-conditional, continuously evolving world simulation from visual/text conditions.

How it differs from predecessors: Genie 1 proved the feasibility of learning interactive environments from video, Genie 2 scaled to a large foundation model, and Genie 3 pushes real-time performance and generality to new heights.

But there’s an important technical issue worth highlighting here: real-time generation ≠ controllability for control.

24fps frame rate is just one of many metrics. For robotic world models, what matters more includes: temporal consistency, action controllability, long-horizon stability, spatial consistency, object permanence, and more. A high frame rate doesn’t mean it’s ready to serve as a robot training environment.

This distinction is critical and represents one of the core challenges facing interactive world models today.

From this, we can extract a core thesis for the entire article: the real technical barrier for world models is not “can they generate the future,” but “can they reliably predict the future given actions.” This thesis will run through the evaluation framework and trend discussions that follow — it connects action controllability, counterfactual prediction, long-horizon stability, and downstream control.

4. Spatial / 3D World Model: World Labs Marble

A Commercial Path for Spatial Intelligence

Fei-Fei Li’s World Labs released Marble in November 2025, generating persistent, downloadable 3D spaces from various media inputs (images, video, text).

Marble and Cosmos both involve world modeling, but their product goals and technical focuses differ significantly. Cosmos leans more toward Physical AI / simulation / synthetic data / robotics, while Marble leans more toward 3D world generation / spatial intelligence / world reconstruction / content creation. Both are advancing toward “generable, interactive, spatially consistent world representations,” but in different directions.

This is an important signal of world models moving toward commercialization — the spatial intelligence path is closer to content creation and AR/VR applications.

What’s even more noteworthy is that World Labs in 2026 has moved well beyond using Marble for spatial content creation. Through acquiring SceniX and launching Real-to-Sim-to-Real (R2S2R), they are actively pushing Spatial World Models toward robot training and evaluation. They demonstrated a complete closed loop from real-to-sim to sim-to-real, including policy training, policy evaluation, zero-real-data training, and simulation ranking vs hardware ranking comparisons. In other words, Spatial World Models are moving from “generating a world that looks real” to “generating a world where robots can learn and be evaluated.”

5. Predictive Representation Learning: AMI Labs and JEPA

LeCun’s Technology Bet

Yann LeCun founded AMI Labs (Advanced Machine Intelligence Labs) in Paris, completing approximately $1.03 billion (approximately €890 million) in funding in March 2026, making it one of the most watched new companies in European AI foundation models in recent years.

AMI Labs’s technical approach is based on the JEPA (Joint Embedding Predictive Architecture) that LeCun has championed for years. JEPA’s core idea is: predict representations rather than raw observations — make predictions in abstract representation space, not pixel space.

It’s important to be precise here: JEPA is more accurately a family of predictive representation learning architectures and ideas, not a complete world-model definition. AMI’s approach is to use this kind of predictive representation as a foundation to learn models that can capture the laws of the real world. The path from JEPA to latent prediction to world model to AGI still contains substantial open research questions at every step.

This shares some common ground with the Dreamer series — as I discussed in the RSSM deep dive, RSSM also makes dynamic predictions in latent space rather than directly predicting the next image frame. But the two are not at the same level: RSSM is a latent dynamics model aimed at supporting planning and control; JEPA is a predictive representation architecture aimed at learning world state representations useful for tasks.

JEPA’s core argument is not that “pixel reconstruction is wrong.” More precisely: for learning high-level semantics and world state representations, requiring the model to precisely predict all pixels is not an ideal learning objective — because pixel space contains a large amount of task-irrelevant details and randomness.

Why pay attention? LeCun is one of the founding figures of deep learning, and his judgment on technical directions carries significant weight. AMI’s research direction continues LeCun’s long-advocated JEPA / predictive representation approach, aiming to build world models that can learn the laws of the real world. This at least indicates that investors and the founding team have a very strong conviction about the future industrial value of this technology path. Of course, whether it produces results depends on execution.

World Labs vs AMI Labs: A Comparison of Two Paths

World Labs and AMI Labs form the most interesting contrast in this article — they represent two extreme directions of world model divergence:

These two paths are not in conflict, but they answer completely different questions.

6. Key Survey Papers: Building a Global Perspective

If you want to read a few papers to build a comprehensive understanding of world models, I recommend these:

“A Definition and Roadmap for World Models”

This preprint (2607.06401) does something very valuable: it provides a formal definition of world models and draws a technology roadmap. The four technology paths above draw primarily from this paper.

“World Model for Robot Learning: A Comprehensive Survey”

This paper focuses specifically on world models in robot learning. If your question is “how do I use world models on real robots?”, this is more targeted. It systematically covers how world models apply to perception, planning, and control.

“A Comprehensive Survey on World Models for Embodied AI”

This paper analyzes world models from the embodied AI perspective. The survey was first published in October 2025 and has been updated to v3 as of June 2026. A maintained paper list is available on GitHub for extended reading.

7. A Comparison Table

Placing the representative work discussed above side by side:

This table is more valuable than paragraphs of adjectives. When you encounter a new “world model,” place it in this framework first to quickly understand its relationship to other work.

8. The Biggest Gap in World Models: Evaluation

After discussing all these projects and technologies, there’s a question we can’t avoid: how do we prove that a world model is “good”?

This is the weakest link in the current world model field, and the dimension I pay closest attention to when reading papers. Multiple surveys from 2025-2026 have listed benchmarks, metrics, physical consistency, computational efficiency, and long-horizon consistency as core open problems.

I believe world model evaluation can be organized as a ladder. The following is not a recognized standard benchmark hierarchy, but rather an evaluation framework I use when reading related work — “from surface capabilities to actual value”:

Generation quality
       ↓
Temporal consistency
       ↓
Physical consistency
       ↓
Action controllability
       ↓
Counterfactual accuracy
       ↓
Long-horizon stability
       ↓
Downstream task improvement

The further down you go, the closer evaluation gets to the world model’s ultimate value in real systems.

Most current work remains at the upper layers — video generation models are impressive at generation quality, but can they achieve physical consistency? Can they be precisely controlled by actions? Do they actually help downstream tasks? These questions often go unanswered.

This is why I’ve repeatedly emphasized “real-time generation ≠ controllability” and “you can’t call something VLA + world model just because it has predictive ability.” Evaluation criteria are shifting from “does it look realistic?” to “are predictions accurate, can it be controlled by actions, and is it actually useful for downstream tasks” — yet most projects are still proving themselves with upper-layer metrics.

What truly deserves attention is work that reaches the bottom of this ladder and demonstrates world model value through downstream task improvement.

9. On VLA + World Model Convergence

I discussed in my VLA vs World Models article that VLA and world models are not competing approaches but complementary ones. In 2026, we are indeed seeing more work combining the two, but this requires particular caution.

You can’t simply label a robot system as “VLA + world model” just because it has prediction or planning capabilities. You need to specify: which module is the world model? Is it an explicit dynamics model, or latent prediction within the policy? Is it training-time simulation, or inference-time planning?

Looking at the 2026 robot learning survey, the most interesting roles for world models within VLA frameworks are not about “stitching together” two systems, but rather world models beginning to take on three specific functions: simulator (generating training experience in imagination), evaluator (assessing policies without real interaction), and data generator (providing synthetic data for VLA post-training). This division of labor is more technically valuable than vaguely saying “VLA + world model.”

If this convergence is truly achieved — with world model latent state representations directly serving as conditional inputs for the VLA, and the VLA’s language grounding ability guiding the world model’s imagination direction — it would be a very powerful architecture. But most current work is still in the exploration phase and requires more specific technical validation.

10. Technology Trends: Three Important Shifts

Trend 1: From Latent Dynamics to Foundation-Scale World Models

The most visible technical evolution direction for world models in 2026 is from single-task small-scale latent dynamics models toward foundation-scale world models. This shift manifests across multiple dimensions: RNN/GRU → Transformer, small latent state → tokenized/spatial representation, single-task dynamics → general-purpose foundation model, short horizon → long horizon, offline prediction → controllable simulation.

I discussed the role of Transformers in world models in a previous article. But the more fundamental change is not “Transformers replaced GRUs” — it’s that world models themselves are evolving from “a dynamic predictor for small environments” into “a foundation model that can generalize to multiple environments.” Both Cosmos and Genie 3 are products of this direction.

Trend 2: Evaluation Criteria Moving from Visual Quality to Downstream Utility

In 2024-2025, the primary question for world models was “can it work?” — can it learn environment dynamics, can it generate useful data through imagination?

In 2026, the question has become “how to make it work well” — how to improve sample efficiency, how to reduce training costs, how to deploy stably on real robots.

In other words, the evaluation criteria for world models are shifting from “does it generate realistically?” to “are predictions accurate, can it be controlled by actions, and is it actually useful for downstream tasks?” The real watershed has moved from generation quality to controllability, predictive accuracy, and downstream utility.

This is why industrial platforms like Cosmos are emerging, and why DreamerV3 training engineering practices (which I covered in detail in this article) have become as important as the algorithms themselves.

Trend 3: World Models Evolving from “Predictors” to “Simulators”

This theme recurs throughout the article, but it deserves to be called out explicitly. World Labs’s own technical framework explicitly identifies the Simulator as the key role connecting world → agent → action → learning → evaluation. The robot world model survey similarly places learned simulator, policy learning, evaluation, and data generation at the core.

When Cosmos generates synthetic data, it is essentially a simulator. World Labs’s R2S2R is essentially a simulator. Genie 3’s interactive world is essentially still a simulator. World models are evolving from “predictors that forecast the future given history” into “simulators where agents can learn, train, and be evaluated.” This is the most noteworthy trend of 2026.

11. Will World Models Ultimately Become Simulators?

Not necessarily.

I’ve made a strong case for simulators above, but an important distinction is needed: the simulator is one of the most important forms of world model application, but not the only form.

Specifically: Dreamer’s world model primarily serves planning; JEPA primarily serves representation learning; Cosmos primarily serves data generation; Genie primarily serves interactive simulation; Marble is pushing spatial representation toward simulator. The ultimate form of each path is not the same.

So a more precise conclusion is: the simulator is one of the most important paths for world models to reach downstream utility, but different world model paths will demonstrate their value in different ways. This echoes the article’s core thesis — “world model” is losing its singular meaning, and the ways it proves its value are equally not singular.

12. Reading Recommendations

Finally, a practical reading framework.

When reading any world model paper, ask 6 questions first:

1. What is the state representation?
2. What does the dynamics model predict?
3. Does action enter the dynamics?
4. How long is the prediction horizon?
5. How do they verify the predictions are actually useful?
6. What is the final downstream task?

Questions 5 and 6 are especially critical. Otherwise you risk the common pattern: video prediction benchmarks look impressive, but they don’t help with robot control. This is precisely one of the most important issues to discuss in the current world model field.

If you’re new to world models:

Start with any one of the surveys mentioned above to build a global perspective. Then work through this blog’s foundational series — from What is a World Model to the RSSM Deep Dive to Dreamer Explained — to solidify core concepts.

If you’re already doing world model research:

Focus on the Cosmos technical report and Genie 3 paper to understand how industry approaches large-scale world models. Then follow AMI Labs’s progress — if the JEPA approach succeeds, it could reshape the field’s technical paradigm.

If you care about engineering deployment:

Cosmos’s open-source toolchain is priority one. Then look at the Sim-to-Real chapters in the World Model for Robot Learning survey. Finally, follow NVIDIA and Google’s latest sharing on robot deployment.

The most noteworthy thing in the world model space in 2026 is not which world model won, but that the term “world model” itself is losing its singular meaning.

Some models focus on generation, some on simulation, some on prediction, some on planning, and some on building spatial representations. The truly important question has shifted from “who is the real world model” to “what role does this model play in the world → state → action → consequence loop?” Understanding this divergence matters more than chasing any single project.

In the next article, I’ll discuss what kind of team you need to start a company in embodied AI — not a generic business plan, but a pragmatic analysis from an engineer’s perspective. Stay tuned.