Tag
LeWAM 提出一种端到端的 JEPA 控制框架，通过联合训练下一隐状态预测与动作流匹配目标，并对隐状态施加各向异性高斯约束，使单一模型同时支持直接策略执行与高效规划，在接触丰富任务和长程任务上将成功率从 28.6% 提升至 89.7% 和从 10.9% 提升至 46.2%，规划速度提升 32.3 倍。
提出了 WorldGuide，一种面向长程程序化任务的闭环视频世界模型，通过规划器与执行器耦合及分层视觉记忆，在自建的 WorldGuide-Bench 上以 33.33% 任务成功率超过 MiniMax-H3。
DreamTrue 是一个多视角、跨具身的机器人世界模型，通过离线几何校准与反事实后训练提升动作跟随与物理合理性，在 AgiBot 数据集上将交互缺陷率从 48.12% 降至 6.25%，并获得 AgiBot World Challenge 2026 世界模型赛道第一名。
RoboJEPA 基于 JEPA 架构，在涵盖 12 种机器人具身的大型真实机器人数据集上训练，首次为多具身机器人世界模型建立了 scaling law：其 latent rollout 的“想象力误差”随计算量呈二阶幂律变化，并可作为真实机器人规划性能的可靠代理。该模型参数量达 8B，是迄今最大的 JEPA 预测模型，支持零样本部署为长时程规划的机器人智能体，并开源全部检查点与训练/部署代码。
A solo researcher shares Part 2 of a local AI world model: a ~960M-parameter transformer using diffusion forcing and text cross-attention that runs at 30-60 FPS on an RTX 5090, accepts live WASD actions, and supports mid-rollout text prompt switching (e.g., 'add a pond to the desert').
Mirror Particle, a two-year-old San Francisco startup, is building a 'world model' foundation model that simulates how and why human behavior changes over time, positioning it as a more fundamental approach than LLM-based role-playing for consumer prediction.
SPW-Nav 是一种流式全景世界模型，能根据语言指令从单张全景图实时生成一分钟2K 360度视频，并支持中途切换指令，用于语言引导的导航与3D场景探索。
Andrej Karpathy shares a simple eval where an LLM is asked "Land or Water?" for latitude/longitude coordinates 16,200 times; the models perform well, suggesting they internalize world knowledge just from compressing the internet.
UniWAM 提出了一种统一的架构，将物理推理器、世界生成器和动作预测器整合在一起，联合学习物理世界的语义理解、视觉生成与动作预测。该模型在多项评估中达到 SOTA，并揭示了统一人机协同训练的对数线性缩放规律。
Lookahead-R 将工具检索重构为受资源约束的序贯决策问题，用轻量级执行感知代理世界模型驱动成本敏感的 MCTS，在 ToolBench 的 I3 分割上取得 91.40% 的 NDCG@5，超过 ToolGen (90.16%)。
Dream4ACT introduces a shared visual action interface ("action views") rendered via URDF-based forward kinematics to enable joint video-action modeling across multiple robot embodiments, combining observations and actions in a single diffusion world model with masked flow matching. It achieves 88.98% success on RoboTwin 2.0 and 65.66 on TriWorldBench using a training-free URDF-constrained recovery mechanism for executable actions.
Microsoft Research introduces Quine, an early-stage AI system that creates a multimodal world model of biology to help scientists computationally explore complex hypotheses and prioritize experiments.
HelixWorld is a real-time interactive audio-visual world model where visual scenes and camera-grounded spatial stereo sound co-evolve under user interaction, using a teacher-student distillation pipeline to run at 24 FPS on a single GPU. The paper also introduces HelixBench to evaluate spatial-acoustic consistency of synthesized sound fields.
MM-ABC提出一个面向通用移动操作的基础模型，通过稀疏多层VLM特征实现空间感知、以世界想象与几何意图为监督的训练分支，以及MM-APT中用掩码联合注意力协调操作与移动流。模型在5000+小时、17种本体的数据上预训练，在RoboCasa365、LIBERO等基准及真实移动操作任务上取得强性能。
EditWorld is a video world model for precise editing and flexible referencing in interactable worlds, achieving the best overall score of 73.8 on the WBench-Editing benchmark.
stable-worldmodel is an open-source platform for standardized and reproducible world modeling research, accepted at NeurIPS 2026, providing a high-performance data layer, baselines, and evaluation suites to reduce research overhead.
Odyssey introduces Agora-2, a next-generation multi-agent world model that supports real-time simulation of up to 20 humans and agents interacting in a shared environment, with a multiplayer research preview now available.
OdysseyML introduces Agora-2, a next-generation multi-agent world model that supports up to 20 humans and agents interacting in real-time within a shared simulated environment, with a multiplayer research preview now available.
PixVerse R2 is a real-time world model that generates continuously evolving audiovisual worlds, accepting text, images, audio, and actions with session memory for interactive and controllable experiences.
The paper introduces THAW-VLA, a method that distills world-model representations into Vision-Language-Action models for robotics, enhancing robustness and performance on simulation and real hardware without additional inference overhead.