World Models

A video generator may produce one physically plausible future without capturing the range of outcomes that could follow the same observation and action. PAWBench evaluates whether repeated rollouts recover that underlying behavioral distribution.

GameWAM combines a visual world model with an action policy for closed-loop gameplay and GUI control. It predicts future observations, generates executable keyboard-mouse trajectories, and studies a new source of bias in generative actions.

A new paper argues that scaling world models requires more than larger collections of web video. Game engines could provide executable environments, grounded rewards, and long-horizon trajectories for post-training spatial models.

Code World Model separates persistent world evolution from visual rendering. A coding agent maintains executable state and consequences, while a video model turns those constraints into visual observations.

EchoWM is an omnimodal world model designed for interactive, enterable media. As users continuously navigate a scene, it jointly generates 720p video, environmental sound, music, and speech while following first-person or third-person camera behavior.

ReWorld separates short-horizon action control from long-term scene memory during training, then combines a bounded KV cache with pose-indexed landmark retrieval at inference. The approach is designed to preserve revisited scenes without allowing context and compute costs to grow without bound.

A new capability-based survey measures world models against the functions of traditional simulators. It finds meaningful progress in interaction and control, but major gaps in physics guarantees, state feedback, and reproducible long-horizon evolution.

ForgeWM introduces a progressive causal training recipe for converting a bidirectional action-conditioned video generator into interactive world models that run in one, two, or four denoising steps. The design targets low-latency gameplay while preserving alignment between controls and generated motion.

WorldRover introduces an Unreal Engine-based pipeline that renders long exploration routes together with geometry, motion, and correspondence annotations. The resulting data is designed for models that must maintain and revisit coherent representations of interactive worlds.

StateFlow proposes a state-centric framework for generative previsualization across film, games, architecture, and urban design. Instead of producing a video in one shot, it maintains an editable 3D world that can be constructed, evolved, and accessed over time.

Video world models can look plausible frame by frame yet drift over long horizons. WorldCycle turns reversible action cycles into a self-verifiable training signal, so “getting back to the starting state” becomes the objective.

MiniWorld aims to make video world model research easier to reproduce by offering an end-to-end training framework that does not rely on adapting large pretrained video generators. Its focus is causal streaming generation, modest compute, and open implementation.

The paper reframes world models as agent-centric interactive proxies, shifting the focus from predicting physical states to delivering usable feedback for planning, learning, and continual improvement.

WCM targets a core bottleneck in robotic VLA post-training: critics that rely on single-frame observations struggle under partial observability. By jointly predicting future latent states and values, it trains the critic to encode temporal dynamics rather than only regress scalar returns.

Roomer reframes residual errors in generated indoor layouts as object-grounded repair tasks, identifying the items responsible for collisions, blocked openings, out-of-bounds placements, and unusable circulation.

WorldExam introduces a hierarchical benchmark for controllable video world models, shifting evaluation from visual appearance and instruction following toward whether a generated world can react coherently to its own state.

ODEWorld moves beyond fixed-step prediction by learning a continuous latent velocity field governed by physical time. The approach targets long-horizon visual prediction, flexible temporal resolution, backward prediction, and planning-friendly dynamics for robotics.

Mental World Modeling argues that a world model can correctly reconstruct a physical scene and still fail at predicting human action unless it tracks hidden mental states.

QQWorld targets a subtle but important weakness in latent world models: heavy-tailed latent distributions can remain insufficiently controlled under existing Gaussian regularizers. The paper replaces the Epps-Pulley objective with quantile-quantile matching to keep corrective signals active in the tails.

World Labs’ acquisition of robotics simulation company SceniX suggests that world models are moving beyond visually coherent 3D generation. The next challenge is to create virtual environments where robots can act, fail, learn and return to the real world.

ShadowDancer targets a central bottleneck in interactive video world models: transferring an action demonstrated in one visual setting to another. Its key idea is to learn dynamics from paired videos that share motion but differ in appearance.