World Models

PhiZero introduces a physical world model built on a compact discrete representation of world-state transitions. Instead of predicting future video purely in pixel space, it first reasons over a learned physical-language sequence and then renders the result.

World Labs has introduced R2S2R, a real-to-sim-to-real engine that reconstructs real robot tasks in simulation, trains and evaluates policies there, and then deploys them back to physical robots.

Wonder reframes video generation as an interactive world model: starting from an image or video, users can move the camera, reveal unseen areas, and revisit earlier views in real time.

NVIDIA has introduced Cosmos-H-Dreams, a real-time, action-conditioned generative simulator for surgical robotics. The system distills Cosmos-H-Surgical-Simulator into a causal student model and serves it with the FlashDreams streaming inference stack.

WorldWeaver introduces world state registers into streaming autoregressive diffusion, giving multi-agent video generation an explicit way to preserve shared environment state across agents and views.

UniWorld-View, developed by Tuzhan Intelligence with Peking University and Pengcheng Laboratory, has reached the top of the WorldScore ranking associated with Fei-Fei Li’s team. The model can generate camera-controlled novel-view videos from a single image or monocular video.

AlayaWorld aims to turn text, images, or videos into explorable and evolving virtual environments. Its technical focus is not only visual quality, but also interaction, long-horizon consistency, and efficient response.

ABot-World-0 frames video generation as an action-conditioned world model that can run in a closed interactive loop. Its main contribution is not only visual synthesis, but the combination of data curation, long-horizon distillation, and streaming inference on a single RTX 5090.

Masked Visual Actions introduces a pixel-space action interface for video models, representing control as a partially revealed trajectory in a video. The same model can predict scene responses to robot motion or infer robot behavior from desired object motion.

AlayaRenderer-Flash accelerates generative world rendering from 0.56 FPS to 31.54 FPS, making a physics-driven generative scene playable at around 30 FPS. Instead of inventing dynamics from prompts, it renders RGB frames from structured world states.

EvolvingWorld moves interactive literary simulation beyond static persona imitation. It treats story generation as a long-horizon process in which characters, scenes, and world state keep changing together.

The paper argues that masked diffusion language models can serve as stronger and more steerable text-based world models for agent training. By denoising bidirectionally, MDLMs better preserve global anchors such as tool schemas, prior turns, rules, and intended outcomes.

NVIDIA has released Cosmos 3 Edge on Hugging Face, a 4B-parameter open world model for robotics and vision AI. The model is designed to combine scene understanding, prediction, simulation, and action generation on edge hardware.

DSWorld applies the world model idea to autonomous data science agents, predicting the outcome of workflow operations before costly execution. The paper reports around 14× faster RL-based training and 3–6× faster search-based inference.

UniVR aims to train models to reason, model physical dynamics, and plan over long horizons directly from visual demonstrations. Its key components are the VR-GRPO reinforcement learning scheme and the VR-X benchmark for purely visual evaluation.

A new arXiv paper reframes the race toward generative game engines around a classic loop: action, state, and observation. Its central message is that playable worlds need persistent rules and evolving states, not just convincing video prediction.

A new arXiv paper introduces M⁴World, a driving world model that generates surround-view video and synchronized LiDAR while supporting object-level manipulation. The work focuses on controllability, long-horizon stability, and rare-case customization for autonomous-driving simulation.