World Models

Tencent ARC Lab introduces Multics, a multi-player video generation framework that organizes communication at player, group, and global levels. The design aims to preserve cross-view consistency while reducing the cost and interference caused by all-to-all communication.

RSI-Anything uses predictive world models to help AI systems choose actions, select learning data, and retain useful policy updates. Its evaluations span code repair, GUI tasks, laboratory planning, reaction optimization, and immune science.

Huawei’s Memento 3 completed all 25 games and 183 levels in the ARC-AGI-3 public set without fine-tuning or changing the base model’s weights. Its advantage comes from turning discovered rules into executable world models that can be tested and revised.

Two years after a dispute with ByteDance over access to training infrastructure, former intern researcher Tian Keyu has reportedly raised nearly $30 million for a stealth startup valued at $200 million. The company is focused on world models but has not revealed a name or product.

DreamTrue is a multi-view, cross-embodiment robot world model designed to predict futures that better reflect commanded actions and physical interactions. It combines geometric calibration, counterfactual post-training, and a human-guided video reward model.

WorldGuide frames procedural video generation as a closed-loop execution problem. It plans an atomic action from the current visual state, renders the action as a video clip, and uses the result to decide what to do next or whether the task is complete.

Trace2Env reconstructs interactive environments from historical traces instead of rebuilding the original executable system. A world model agent then simulates observations and persistent state changes for task agents.

WorldSonus is an interactive video-to-audio framework for world models, designed to generate controllable spatial stereo sound in real time. The paper reports an RTF of 0.41 and strong generalization to open-domain video-to-audio benchmarks.

RoboJEPA uses real-robot data from 12 embodiments to study how latent world models improve with scale. Its latent rollout error tracks downstream planning quality and can help predict performance beyond the tested training scale.

A new benchmark evaluates video world models by measuring observable physical relationships rather than relying only on visual plausibility or reference videos. Tests across 1,280 videos show that convincing appearance still falls short of physical reliability.

HLA-WM combines geometry-guided retrieval with recurrent linear attention to recover relevant history during long video rollouts. It improves scene revisits and camera control without additional training while sharply reducing memory use.

RealtimeWAM tackles two major latency sources in world action models: iterative action denoising and sequential execution between video and action experts. The paper reports less than 1% accuracy degradation and up to 25× end-to-end speedup on H100.

Video world models rely on growing KV caches to maintain long-range consistency, but aggressive quantization can cause flicker. QuantWM addresses the problem by preserving attention behavior rather than minimizing reconstruction error alone.

JEPA-TTT enables a pretrained JEPA world model to keep adapting when test-time dynamics differ from training. It updates only the latent dynamics predictor through persistent self-supervision, while preserving the visual representation and task objective.

ProWAM extends world action models with an ordered sequence of sparse visual subgoals, giving robots intermediate guidance without generating dense future video. The reported results show stronger robustness on several simulation benchmarks.

HelixWorld extends interactive world models beyond silent visual rendering by jointly generating scenes and camera-grounded stereo audio. The work combines spatially aligned data, teacher–student distillation, and a benchmark for evaluating acoustic consistency.

The World Embedding Benchmark evaluates whether video embeddings align with physical concepts and preserve quantitative information. Its results reveal a gap between having physical signals in a representation and making those signals usable across modalities.

World Observer separates acting from observing by jointly generating an agent-centric actor view and one or more panoramic observer views. This lets out-of-view objects continue evolving and improves their state and appearance when they re-enter the actor’s view.

AutoGUIWorld combines an image generator with a task planner to synthesize GUI interaction trajectories without deploying the underlying software. The resulting data improves computer-use agents across several desktop and browser environments.

LOCI combines key-value caching with recurrent linear memory to help video world models reproduce previously observed regions when a camera returns. Its geometry-aware memory improves scene retrieval while reducing the cost of processing long videos.

NVIDIA researchers introduce Physis-Lang, a framework that uses detailed physical language as a shared representation for video curation, training, and generation. Assertion-level evaluation and language-guided retrieval help models produce videos with more plausible physical behavior.