顶尖研究者预测 2027 年世界模型将迎突破
World Models Breakthrough in 2027 – Top Researchers
多位顶尖研究者按子领域梳理了世界模型的研究版图，涵盖 Yann LeCun 的 JEPA、Fei-Fei Li 的 3D 空间智能、Danijar Hafner 的 Dreamer 系列、Tim Rocktäschel 的 Genie 等方向。Danijar Hafner 的 Dreamer 4（2025）成为首个仅靠离线数据、无需环境交互即在 Minecraft 中获得钻石的智能体。
What is a world model?
A world model is an AI system that learns an internal, predictive representation of how an environment evolves — letting an agent imagine future states, plan actions, and reason about cause and effect without acting in the real world first. World models span video generation, interactive simulation, 3D spatial intelligence, JEPA-style joint-embedding prediction, and model-based reinforcement learning. These predictive models provide training, simulation, and planning infrastructure for future agents and robots. The investment surface spans both model architectures and the deployment stack built around them.
World model researchers by subfield
| Subfield | Leading researchers | Why it matters | Notable work |
|---|---|---|---|
| Self-supervised world models (JEPA) | Yann LeCun · Mahmoud Assran · Michael Rabbat · Saining Xie · Pascal Vincent · Ishan Misra · Piotr Bojanowski | Learn predictive representations from video without generating pixels. | A Path Towards Autonomous Machine Intelligence (JEPA) |
| Video generation as world simulation | Bill Peebles · Tim Brooks · Sergei Tulyakov · Cristóbal Valenzuela · Agrim Gupta · Zangwei Zheng | Large-scale video models that double as simulators of the visual world. | Scalable Diffusion Models with Transformers (DiT) |
| 3D & spatial intelligence | Fei-Fei Li · Justin Johnson · Ben Mildenhall · Christoph Lassner · Aleksander Holynski · Jiajun Wu | Persistent, navigable 3D worlds reconstructed or generated from images and text. | NeRF: Representing Scenes as Neural Radiance Fields |
| Model-based RL & agents | Danijar Hafner · David Ha · Nicklas Hansen · Jurgis Pasukonis · Pieter Abbeel · Eloi Alonso | Agents that learn and plan inside a learned, imagined world model. | Mastering Diverse Domains through World Models (DreamerV3) |
| Interactive & playable world models | Tim Rocktäschel · Jack Parker-Holder · Stephen Spencer | Generate controllable, playable environments from a single prompt. | Genie: Generative Interactive Environments |
| Driving & navigation world models | Alex Kendall · Anthony Hu · Amir Bar · Gaoyue Zhou · Yunzhu Li | Predictive world models for autonomous driving and visual navigation. | Learning to Drive in a Day |
PioneersField-definers · 8
Pieter Abbeel
AcademiaFrontier lab
Professor of EECS, UC Berkeley · Amazon Scholar / Head of LLM efforts in AGI org, Amazon
PhD · Stanford University — Andrew Ng
Berkeley deep-RL/robot-learning leader; co-founded Covariant; now heads Amazon's LLM/AGI robotics efforts
Pioneer of deep RL and imitation learning. Drove UniSim and Large World Model, and recent generalist robot world models (DreamDojo) trained on tens of thousands of hours of human video.
LinkedInWebsiteScholarWikipediaXCovariantCovariant on X
Key papers
Chelsea Finn
AcademiaStartup
Associate Professor of Computer Science & EE, Stanford University · Co-Founder, Physical Intelligence
PhD · UC Berkeley — Sergey Levine and Pieter Abbeel
Co-creator of MAML meta-learning; co-founder of Physical Intelligence (robot foundation models)
Pioneer of meta-learning and visual foresight for robots. Co-author of Physical Intelligence's pi-0.6 (RL self-improvement for VLAs) and Ctrl-World / VLAW, which use video world models to evaluate and improve robot policies.
LinkedInWebsiteScholarWikipediaXPhysical IntelligencePhysical Intelligence on X
David Ha
Startup
Co-founder & CEO, Sakana AI
PhD · University of Tokyo
Author of the seminal 'World Models' paper; co-founder/CEO of Sakana AI (nature-inspired foundation models)
Co-author of the foundational 2018 "World Models" paper that introduced the modern framing of learning compact predictive models of an environment for control. Now at Sakana AI.
LinkedInWebsiteScholarXSakana AISakana AI on X
Danijar Hafner
Startup
Co-founder, Embo
PhD · University of Toronto — Advised by Jimmy Ba (Vector Institute / U. Toronto)
Creator of the Dreamer line of world-model agents (DreamerV3, Dreamer 4); RL by imagination
Creator of the Dreamer line of model-based RL. Dreamer 4 (2025) was the first agent to obtain diamonds in Minecraft purely from offline data with no environment interaction — a landmark for training agents inside imagined world models.
Key papers
- Mastering Diverse Domains through World Models (DreamerV3)
- Training Agents Inside of Scalable World Models (Dreamer 4)
Kaiming He
AcademiaFrontier lab
Associate Professor of EECS (tenured), MIT · Distinguished Scientist (part-time), Google DeepMind
PhD · The Chinese University of Hong Kong — Xiaoou Tang
Lead author of ResNet (deep residual learning); also Faster R-CNN, Mask R-CNN, MAE
Author of ResNet and Masked Autoencoders (MAE), among the most-cited works in deep learning. Recent generative work (JiT, improved Mean Flows) pushes diffusion-style models toward single-step, tokenizer-free generation — core infrastructure for video world models.
Yann LeCun
StartupAcademia
Founder & Executive Chairman, AMI Labs (Advanced Machine Intelligence) · Professor, New York University
PhD · Université Pierre et Marie Curie (Sorbonne / UPMC), Paris — PhD on backpropagation; later postdoc with Geoffrey Hinton at U. Toronto
Turing Award pioneer of CNNs/deep learning; champion of JEPA world-model architectures
Turing Award winner and originator of the Joint-Embedding Predictive Architecture (JEPA), the leading self-supervised alternative to autoregressive world models. His team's 2025–26 work extended JEPA into a deployable family — VL-JEPA, Causal-JEPA, and latent-action world models learned from in-the-wild video.
LinkedInWebsiteScholarWikipediaXAMI Labs (Advanced Machine Intelligence) on X
Key papers
- A Path Towards Autonomous Machine Intelligence (JEPA position paper)
- Gradient-Based Learning Applied to Document Recognition (LeNet/CNNs)
Fei-Fei Li
StartupAcademia
Co-founder & CEO, World Labs · Sequoia Professor of Computer Science; Co-Director, HAI, Stanford University
PhD · California Institute of Technology (Caltech) — Advised by Pietro Perona (and Christof Koch)
Created ImageNet; co-founder/CEO of World Labs building spatial-intelligence world models (Marble)
Creator of ImageNet and the intellectual leader of "spatial intelligence." Co-founded World Labs, which raised ~$1B and shipped Marble (persistent 3D worlds from text/image), the World API, and real-time interactive world models running on a single GPU.
LinkedInWebsiteScholarWikipediaXWorld Labs
Key papers
- ImageNet: A Large-Scale Hierarchical Image Database
- ImageNet Large Scale Visual Recognition Challenge
Tim Rocktäschel
StartupAcademia
Co-founder & CEO, Recursive (Recursive Superintelligence) · Professor of Artificial Intelligence, University College London
PhD · University College London (UCL) — Advised by Sebastian Riedel
Led Genie foundation world models & Open-Endedness at DeepMind; UCL AI professor; co-founder of Recursive
Leads interactive world-model research at Google DeepMind, including the Genie line of foundation world models that generate playable, controllable environments from a single prompt.
LinkedInWebsiteXRecursive (Recursive Superintelligence) on X
Key papers
Leading researchersShaping the field today · 30
Mahmoud Assran
Frontier lab
Research Scientist, Meta FAIR
PhD · McGill University — Mila (Quebec AI Institute)
Lead author of I-JEPA, V-JEPA and V-JEPA 2 self-supervised world-model architectures at Meta FAIR
Lead author of I-JEPA and V-JEPA, the image and video instantiations of LeCun's self-supervised world-model paradigm at Meta AI.
Key papers
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA)
- Revisiting Feature Prediction for Learning Visual Representations from Video (V-JEPA)
- V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
Piotr Bojanowski
Frontier lab
Research Scientist & Manager, Meta (FAIR)
PhD · Ecole Normale Superieure (ENS) Paris — WILLOW team; advisors Ivan Laptev, Jean Ponce, Cordelia Schmid, Josef Sivic
FAIR Paris research lead; co-author of fastText, DINO/DINOv2, and I-JEPA self-supervised models
Key papers
- Enriching Word Vectors with Subword Information (fastText)
- DINOv2: Learning Robust Visual Features without Supervision
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA)
Tim Brooks
Frontier lab
Research Scientist, Meta Superintelligence Labs
PhD · UC Berkeley — Alyosha Efros (BAIR)
Co-created OpenAI's Sora; invented InstructPix2Pix; now builds multimodal/world models at Meta
Co-led OpenAI's Sora and now works on Google DeepMind's Veo — the frontier of large-scale video generation as a world simulator.
Key papers
- InstructPix2Pix: Learning to Follow Image Editing Instructions
- Video generation models as world simulators (Sora technical report)
Andrew Dai
Startup
Co-founder & CEO, Elorian
PhD · University of Edinburgh
Co-founder/CEO of Elorian (visual reasoning, world models); ex-DeepMind Gemini data-area lead
Trevor Darrell
Academia
Professor (EECS), UC Berkeley
PhD · MIT — MIT, advised by Eric Grimson
UC Berkeley EECS professor, co-founder/co-lead of BAIR; leading computer vision and deep learning researcher.
Key papers
- Rich feature hierarchies for accurate object detection and semantic segmentation (R-CNN)
- Fully Convolutional Networks for Semantic Segmentation
Agrim Gupta
Frontier lab
Research Scientist, Meta
PhD · Stanford University — Fei-Fei Li
Stanford PhD on video generation (W.A.L.T) and embodied AI; now world/simulation models at Meta
Key papers
- Photorealistic Video Generation with Diffusion Models (W.A.L.T)
- MaskViT: Masked Visual Pre-Training for Video Prediction
Nicklas Hansen
Academia
PhD Candidate (final-year), UC San Diego
Creator of TD-MPC / TD-MPC2 scalable world models for continuous control RL
Author of TD-MPC2 and Newt, scalable model-based RL world models for continuous control across hundreds of tasks.
Key papers
- TD-MPC2: Scalable, Robust World Models for Continuous Control
- Temporal Difference Learning for Model Predictive Control (TD-MPC)
Aleksander Holynski
Frontier labAcademia
Research Scientist, Google DeepMind · Assistant Professor, Columbia University
PhD · University of Washington — GRAIL / UW Graphics (advisors: Steve Seitz, Brian Curless, Rick Szeliski)
Google DeepMind research scientist & Columbia professor; generative video/3D and world-model research.
Anthony Hu
Startup
Researcher, General Intuition
PhD · University of Cambridge — Machine Intelligence Lab (advisor: Roberto Cipolla)
Led GAIA generative world models for autonomous driving at Wayve; now building world models at General Intuition.
LinkedInWebsiteXGeneral Intuition
Key papers
- GAIA-1: A Generative World Model for Autonomous Driving
- GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving
- FIERY: Future Instance Prediction in Bird's-Eye View from Surround Monocular Cameras
Xun Huang
Startup
Founder & CEO, Morpheus
PhD · Cornell University — advised by Serge Belongie
Created AdaIN (now used across diffusion models); founder of Morpheus, real-time interactive video world models
Key papers
- Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization (AdaIN)
- Multimodal Unsupervised Image-to-Image Translation (MUNIT)
Phillip Isola
Academia
Associate Professor of EECS, MIT
PhD · MIT — Ted Adelson (Brain & Cognitive Sciences)
Co-creator of pix2pix and CycleGAN; Platonic Representation Hypothesis on shared world representations
Key papers
- Image-to-Image Translation with Conditional Adversarial Networks (pix2pix)
- The Platonic Representation Hypothesis
Justin Johnson
StartupAcademia
Co-Founder, World Labs · Adjunct Assistant Professor, University of Michigan
PhD · Stanford University — Stanford Vision Lab (advisor: Fei-Fei Li)
Co-founder of World Labs; pioneered real-time neural style transfer and visual reasoning.
Co-founder of World Labs and a leader in 3D generation and visual reasoning; co-author of widely used vision and neural-style-transfer methods.
Lukasz Kaiser
Frontier lab
Member of Technical Staff (Research Scientist), OpenAI
PhD · RWTH Aachen University — Mathematical Foundations of Computer Science (advisor: Erich Grädel)
Co-author of 'Attention Is All You Need' (Transformer); research lead on OpenAI's o1 reasoning models.
Alex Kendall
Startup
Co-founder & CEO, Wayve
PhD · University of Cambridge — Machine Intelligence Laboratory, advised by Roberto Cipolla
Co-founder and CEO of Wayve; pioneer of end-to-end deep learning ('AV2.0') for autonomous driving.
CEO of Wayve and architect of GAIA, generative driving world models for autonomous vehicles.
Key papers
- Learning to Drive in a Day
- Bayesian SegNet: Model Uncertainty in Deep Convolutional Encoder-Decoder Architectures for Scene Understanding
Christoph Lassner
Startup
Co-founder, World Labs
PhD · University of Tübingen — Max Planck Institute for Intelligent Systems / Bernstein Center for Computational Neuroscience, Tübingen
Co-founder of World Labs; computer vision/graphics researcher in 3D/4D reconstruction and neural rendering (Pulsar).
Yunzhu Li
Academia
Assistant Professor of Computer Science, Columbia University
PhD · MIT — Advised by Antonio Torralba and Russ Tedrake (MIT CSAIL)
Structured/particle-based learned world models for robot manipulation of deformable objects and fluids.
Key papers
- Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids (DPI-Net)
- Visual Grounding of Learned Physical Models
Hao Liu
Frontier lab
Research Scientist, Google DeepMind
PhD · UC Berkeley — Pieter Abbeel
Creator of RingAttention and the Large World Model (LWM); 1M-context multimodal video-language world model
Key papers
- World Model on Million-Length Video And Language With Blockwise RingAttention
- Ring Attention with Blockwise Transformers for Near-Infinite Context
Ben Mildenhall
Startup
Co-Founder, World Labs
PhD · UC Berkeley — Berkeley graphics/imaging (advisor: Ren Ng)
Lead author of NeRF (Neural Radiance Fields); co-founder of World Labs.
Inventor of Neural Radiance Fields (NeRF), the technique that launched modern neural 3D reconstruction; co-founder of World Labs.
Ishan Misra
Frontier lab
Director, Research Scientist, Meta Superintelligence Labs (TBD Labs)
PhD · Carnegie Mellon University — Robotics Institute (advised by Martial Hebert / Abhinav Gupta)
Self-supervised vision researcher; co-author of I-JEPA, DINO, and Movie Gen / Emu Video generation
Key papers
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA)
- Emerging Properties in Self-Supervised Vision Transformers (DINO)
- Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning
Jack Parker-Holder
Frontier labAcademia
Research Scientist, Open-Endedness Team, Google DeepMind · Honorary Lecturer, University College London
PhD · University of Oxford — Stephen Roberts (Machine Learning Research Group)
Co-lead of DeepMind's Genie world models (Genie 1/2/3) for generative interactive environments
Genie lead at Google DeepMind, driving interactive foundation world models that generate controllable environments.
Jurgis Pasukonis
Startup
Member of Technical Staff, Project Prometheus
PhD · Queen Mary University of London — Theoretical physics (string theory) PhD; later RL self-taught
Co-author of DreamerV3 world-model RL agent and the Memory Maze long-term-memory benchmark
LinkedInProject Prometheus on X
Key papers
- Mastering Diverse Domains through World Models (DreamerV3)
- Evaluating Long-Term Memory in 3D Mazes (Memory Maze)
Bill Peebles
Frontier lab
Former Head of Sora (departed April 2026), OpenAI
PhD · UC Berkeley — Alyosha Efros (BAIR)
Invented the Diffusion Transformer (DiT); co-creator/Head of OpenAI's Sora
Co-creator of the Diffusion Transformer (DiT) architecture that underpins Sora and most modern video-generation models; now at OpenAI.
Michael Rabbat
Startup
VP of World Models, AMI Labs
PhD · University of Wisconsin-Madison — Robert Nowak (Electrical Engineering)
Co-creator of I-JEPA/V-JEPA self-supervised world models; now VP World Models at LeCun's AMI Labs
LinkedInWebsiteXMeta AI (FAIR)
Key papers
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA)
- Revisiting Feature Prediction for Learning Visual Representations from Video (V-JEPA)
- V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
Sergei Tulyakov
Frontier lab
Director of Research, Creative Vision, Snap Research
PhD · University of Trento
Director of Research at Snap; video generation pioneer (MoCoGAN, First Order Motion Model, Snap Video)
Key papers
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis
- MoCoGAN: Decomposing Motion and Content for Video Generation
- First Order Motion Model for Image Animation
Cristóbal Valenzuela
Startup
Co-Founder & CEO, Runway
Co-founder & CEO of Runway; generative video moving into general world models (GWM-1)
CEO of Runway, which ships commercial generative world models for video (Gen-series, GWM-1) used widely in film and media production.
LinkedInWebsiteWikipediaXRunway
Pascal Vincent
Frontier labAcademia
Research Scientist, Meta (FAIR) · Adjunct Professor (Mila founding member), Universite de Montreal / Mila
PhD · Universite de Montreal — Yoshua Bengio (Mila)
Pioneer of denoising autoencoders & denoising score matching; co-author of I-JEPA self-supervised models
Key papers
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA)
- Stacked Denoising Autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion
Jiajun Wu
Academia
Assistant Professor of Computer Science (courtesy in Psychology), Stanford University
PhD · MIT — William T. Freeman and Joshua B. Tenenbaum
Physical scene understanding; physics-engine-integrated and neuro-symbolic visual reasoning
Leads physics-based and 3D world-model research at Stanford, including multi-scale 3D world generation (WonderZoom) and intuitive-physics models.
Key papers
- The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision
- Galileo: Perceiving Physical Object Properties by Integrating a Physics Engine with Deep Learning
Zhou Xian
Startup
Co-founder & CEO, Genesis AI
PhD · Carnegie Mellon University — Robotics Institute, advised by Katerina Fragkiadaki
Co-founder/CEO of Genesis AI; creator of RoboGen and the Genesis generative physics engine for robotics.
Key papers
- RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation
- Towards Generalist Robots: A Promising Paradigm via Generative Simulation
Saining Xie
StartupAcademia
Co-founder & Chief Science Officer, AMI Labs · Assistant Professor of Computer Science (on leave), New York University
PhD · UC San Diego — Zhuowen Tu
Co-creator of Diffusion Transformers (DiT) and ConvNeXt; co-founder/CSO of AMI Labs (JEPA world models)
Co-creator of V-JEPA and a leader in generative representation learning; recent work on flow-map distillation and representation autoencoders set new image-generation SOTA.
LinkedInWebsiteXAMI Labs (Advanced Machine Intelligence)AMI Labs (Advanced Machine Intelligence) on X
Sherry Yang
AcademiaFrontier lab
Assistant Professor of Computer Science, NYU Courant · Staff Research Scientist, Google DeepMind
PhD · UC Berkeley — Pieter Abbeel
Generative world models / real-world simulators (UniSim) for embodied agents
Pioneer of treating video generation as a universal world model and simulator; co-author on Gemini Robotics and Veo-based policy evaluation at Google DeepMind.
Rising starsEmerging first-authors · 8
Eloi Alonso
Startup
Co-Founder, General Intuition
PhD · University of Geneva — François Fleuret's group
Led DIAMOND and IRIS diffusion/transformer world models; co-founder of General Intuition
First author of DIAMOND (NeurIPS 2024), a diffusion world model achieving human-level Atari play.
LinkedInWebsiteXGeneral Intuition
Key papers
- Diffusion for World Modeling: Visual Details Matter in Atari (DIAMOND)
- Transformers are Sample-Efficient World Models (IRIS)
Amir Bar
AcademiaFrontier lab
Incoming Assistant Professor, Imperial College London · Research Scientist, Meta FAIR
PhD · Tel Aviv University / UC Berkeley — Advised by Amir Globerson and Trevor Darrell
Navigation World Models (CVPR 2025 Best Paper Honorable Mention); self-supervised visual world modeling
First author of Navigation World Models (CVPR 2025 Honorable Mention), a billion-parameter conditional diffusion transformer for visual navigation prediction.
Boyuan Chen
Academia
Dickinson Family Assistant Professor, Duke University
PhD · Columbia University — Hod Lipson's Creative Machines Lab
Dickinson Family Asst. Prof. at Duke; directs General Robotics Lab; self-modeling robots & automated scientific discovery
First author of Diffusion Forcing (NeurIPS 2024), a training paradigm unifying video prediction and world models; recent work on large video planners for robot control.
Key papers
- Automated discovery of fundamental variables hidden in experimental data (Nature Computational Science)
- Full-body visual self-modeling of robot morphologies (Science Robotics)
Yilun Du
AcademiaFrontier lab
Assistant Professor, Harvard University · Senior Research Scientist, Google DeepMind
PhD · Massachusetts Institute of Technology — Leslie Kaelbling, Tomas Lozano-Perez & Joshua Tenenbaum (MIT EECS)
Harvard SEAS Asst. Prof & Kempner Investigator; compositional/energy-based generative world models
Pioneer of composable diffusion and energy-based world models; frequent collaborator on video-model-as-simulator robotics work.
Key papers
- Learning Interactive Real-World Simulators (UniSim)
- Compositional Visual Generation and Inference with Energy Based Models
Stephen Spencer
Frontier lab
Research Engineer, Google DeepMind
DeepMind researcher; technical lead/contributor on the Genie 2 and Genie 3 foundation world models
Known for Genie 2 (2024). Technical lead on Genie interactive world models
Key papers
- Genie 2: A large-scale foundation world model (DeepMind)
- Genie 3: A new frontier for world models (DeepMind)
Tsun-Hsuan Wang
StartupAcademia
Co-Founder, Genesis AI · PhD Candidate, MIT CSAIL
PhD · Massachusetts Institute of Technology — Daniela Rus (MIT CSAIL Distributed Robotics Lab)
MIT CSAIL roboticist; co-founder of Genesis AI universal robotics foundation model startup
Co-lead on the Genesis physics engine for embodied AI, reported 10–80× faster than prior simulators.
Key papers
- DiffuseBot: Breeding Soft Robots with Physics-Augmented Generative Diffusion Models
- SoftZoo: A Soft Robot Co-design Benchmark For Locomotion In Diverse Environments
Zangwei Zheng
Startup
Video & Inference Lead, HPC-AI Tech
Founder/lead of Open-Sora, the first open-source Sora-like video generation model
Known for Open-Sora (2024). Lead on Open-Sora, open-source video generation framework
LinkedInWebsiteXHPC-AI Tech (Open-Sora)HPC-AI Tech (Open-Sora) on X
Key papers
- Open-Sora: Democratizing Efficient Video Production for All
- Open-Sora 2.0: Training a Commercial-Level Video Generation Model in $200k
Gaoyue Zhou
Academia
PhD Student, New York University
Lead author of DINO-WM world models for zero-shot planning on pre-trained visual features
Known for Navigation World Models (CVPR 2025). Co-author on navigation world models with Amir Bar
Key papers
- DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning
- Navigation World Models
Companies building in world model
- AMI LabsX ↗ — Yann LeCun's lab; JEPA world models predicting in representation space
- Black Forest LabsX ↗ — FLUX visual models unifying perception, generation, reasoning
- DecartX ↗ — Real-time interactive world models at low latency
- ElorianX ↗ — Andrew Dai's multimodal visual reasoning / world model startup
- General IntuitionX ↗ — World models from gameplay video (DIAMOND team); games and the real world
- Genesis AIX ↗ — Genesis World physics simulation + general-purpose robots
- Genie (Google DeepMind)X ↗ — Real-time interactive world model generating playable 3D environments
- GenmoX ↗ — Open video generation models (Mochi)
- HiggsfieldX ↗ — Controllable video generation and motion models
- HPC-AI TechX ↗ — Open-Sora open video generation; Colossal-AI training infrastructure
- Luma AIX ↗ — AI 3D creation and video generation (Dream Machine)
- Moonlake AIX ↗ — Interactive, physics-accurate world models from Stanford team
- Morpheus — Real-time interactive video world models for physical AI
- Niantic SpatialX ↗ — Large geospatial model; visual positioning and 3D maps
- Nucleus4DX ↗ — Gaussian-splatting 3D capture; digital twins and spatial data
- OdysseyX ↗ — Real-time multimodal world models (Starchild-1)
- Reka AIX ↗ — Multimodal foundation models across text, image, video
- RunwayX ↗ — Video generation and general world models (GWM-1)
- SpAItialX ↗ — AI-native 3D foundation models turning images into interactive worlds
- Twelve LabsX ↗ — Video-native multimodal AI for understanding and search
- WayveX ↗ — GAIA driving world models for autonomous vehicles
- World LabsX ↗ — Fei-Fei Li's spatial intelligence lab; Marble 3D world generation
Frequently asked questions
Who is the most influential researcher in world models?
Yann LeCun (AMI Labs / NYU) is widely regarded as the most influential, having defined the JEPA self-supervised paradigm that anchors one of the two dominant approaches to world models. Fei-Fei Li (World Labs / Stanford) and Danijar Hafner (creator of the Dreamer line of model-based RL) are the other two figures most often cited as field-defining.
What is a world model in AI?
A world model is an AI system that learns a predictive internal model of an environment, enabling an agent to imagine future outcomes and plan rather than relying solely on trial-and-error in the real world. Examples include DeepMind's Dreamer, Meta's JEPA, OpenAI's Sora, and World Labs' Marble.
Which companies are leading in world models?
World Labs leads on 3D spatial world models, Google DeepMind on interactive simulation (Genie) and video (Veo), Meta AI on the JEPA architecture, and Physical Intelligence applies world models to robotics.
How is this guide compiled?
This guide maps researchers, their attributable work, subfields and relevant companies from linked public sources. Profiles are grouped as Pioneers, Leading researchers and Rising stars. The technical content is a July 6, 2026 snapshot; X profiles and the public follow list were checked on October 8, 2026. Unverified social accounts are omitted.
More on the radar
来源：Hacker News · changxu.dev