One of the researchers behind Diffusion Forcing reflects on the impact of his seminal paper and the future of world models

Diego Marti Monso, author of the pivotal Diffusion Forcing paper, has been busy. He specializes in world models, which are systems that learn about their surroundings through sensory information and use what they learn to anticipate what may happen next. Currently, world models are seen as the next frontier in AI, offering a promising path toward general robotics, video generation, and virtual reality. For now, though, Marti Monso is focusing on robotics.

His focus reflects his belief that robotics is the next frontier for world models — and, by extension, AI — to conquer. This reflects his belief that generalization and transferable capabilities are the path to success. While he sees value in classical robotics, which are the specialized systems common in warehouses and factories today, Marti Monso expects broader-purpose systems to eventually replace many narrow ones.

Consequently, his research focuses on training methods that remain useful as applications change. "I am interested in the underlying capability of the product," said Marti Monso. "If the method can carry knowledge across different problems, it has a better chance of remaining useful when the application changes."

This philosophy led Marti Monso to Diffusion Forcing, a training method that he co-invented while working as a visiting researcher in the Scene Representation Group at MIT Computer Science and Artificial Intelligence Laboratory. The research, published at NeurIPS in 2024, addressed a problem that arises when AI models predict sequences over an extended period. Previous world models could experience significant drift during long predictions, which made it more difficult to maintain coherent sequences over time.

Diffusion Forcing introduced a training formulation designed to support long sequences while enabling the model to respond to actions as the prediction unfolds.

Marti Monso was one of two researchers who developed the core training formulation, implemented the algorithm, and conducted experiments and evaluations that included video prediction and robotics tasks. Marti Monso never imagined the impact his paper would have on the AI world.

Since then, Diffusion Forcing has moved well beyond MIT’s lab, finding its way into commercial products and open-source projects. Two years after its publication, the work has received over 600 citations and 1,200 GitHub stars, an unusually high number that speaks to its widespread adoption in academia and industry.

For example, Skywork AI, a Shenzhen-listed company with 400 million monthly active users, built SkyReels-V2 directly on Diffusion Forcing. Overworld's Waypoint-1, a real-time, interactive world model with a live consumer product, was also built on Diffusion Forcing. Public materials show that companies such as Google DeepMind, OpenAI, Nvidia, Meta, Adobe, ByteDance, Tencent, and Sony have used or built upon the method. Academic researchers at MIT, Stanford, Harvard, UC Berkeley, and other institutions have also developed workarounds it.

For Marti Monso, this is a point of pride. "A method becomes more useful when people can apply it to problems that it wasn't specifically designed to solve," he said. "That's part of what I look for in research. I want to work on ideas that continue to matter outside the experiment where they began."

This perspective also influences how Marti Monso views the current robotics landscape. Vision-language-action models are a common approach to combining visual information, language-based knowledge, and robot actions. He acknowledges that they currently work well. Nevertheless, he predicts that this approach may eventually prove transitional because he believes that world-model training can develop a richer understanding of physical dynamics. He expects future systems to draw on knowledge across different kinds of data rather than preserving the current boundaries between model categories.

"I think the divisions between modalities will become much less important," said Marti Monso. "Knowledge learned from one kind of experience should be accessible when the system has to reason about something else." He expects future models to benefit from information acquired across multiple forms of data. However, he presents this outcome as a prediction rather than as something the field has already achieved. Robotics remains difficult precisely because strong performance on digital tasks has not cleanly translated into the physical abilities that humans acquire almost automatically.

Marti Monso believes that world models are the next frontier because of this gap. His goal is to build an embodied system that can solve tasks it was never directly trained to perform. This system would be rooted in a generalized understanding of physics and the environment. Such a system could apply its knowledge across changing environments and adapt dynamically, a capability sorely lacking in classical robotics. Monso views the ability to transfer knowledge as a central requirement for moving beyond today’s specialized machines.

The products built around these capabilities may change repeatedly before this goal is achieved. Marti Monso expects hardware and applications to evolve as broader-purpose systems become possible. This uncertainty is precisely why he has chosen to work farther down the stack, where a successful method continues to matter even when the visible technology around it changes. "The application shows what a system can do right now," he said. "I want to work on the methods that expand the range of things it can learn to do next."