Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2610.39721
WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
WorldProbe is a closed-loop benchmark for testing whether video world models preserve a coherent, explorable world during long-horizon camera interaction. A goal-conditioned multimodal agent adapts camera actions online across 164 test cases and four probing families, evaluating 10 video world models over rollouts reaching 180 seconds. The study finds that newer models sustain local visual and content quality longer, but all evaluated systems remain weak at long-range spatial memory and coherent evolution, exposing failures that short, fixed-trajectory benchmarks often miss.
How it works
Each probe begins with an initial frame and a spatial objective. The goal-conditioned multimodal agent receives the task specification, a short recent-frame window, and auxiliary spatial states, then predicts a discrete camera action. That action is converted to the evaluated model's control format, applied to the next generation chunk, and fed back into the next decision. The four task families are geometry-constrained approach, open-ended traversal, cross-view inspection, and successive region exploration.
What the paper evaluates
The study evaluates 10 publicly available video world models using task-specific rollout ranges of 10–60 s, 20–120 s, 20–90 s, and 30–180 s for the four task families. The evaluation uses Qwen3.6-27B as the VLM backbone, Segment-Anything-3.1 for target tracking, and Depth-Anything-3 for local geometry reconstruction. The authors report approximately four days of evaluation on 24 A100 GPUs. For local quality, visual scores are normalized to 0–1 and content scores remain on a 0–10 scale; more negative temporal slopes indicate faster degradation.
Limit to keep in view
The benchmark contains 164 manually constructed cases matched to initial frames from public datasets, so its coverage may depend on the selected scenes, prompts, and spatial objectives.
arXiv:2610.33807
RoboWorM: Evaluating Movement Realism, Consistency and Acceleration Affects in Embodied Video World Models
RoboWorM evaluates whether robotic video world models preserve physical structure as they generate faster. It combines a benchmark of real robot episodes with two deterministic, VLM-free metrics: KIVA measures within-view kinematic stability, and RIGIS measures consistency across camera views. On the authors’ human-alignment evaluations, KIVA reaches 73.5% hallucination-detection accuracy versus 58.4% for frontier VLMs, while RIGIS reaches a Spearman correlation of 0.749 versus 0.326. An audit of three world-model families and nine acceleration methods finds that caching offers the best reported speed-quality balance, including a 2.3× maximum speedup; the results also show that perceptual quality alone can miss structural degradation.
How it works
The benchmark draws on five public teleoperation datasets and covers single-arm, bimanual, and humanoid morphologies, with single- and multi-view camera configurations. Tasks are divided into Markovian and non-Markovian regimes. The paper’s benchmark description reports 1,605 episodes in its overview and 1,608 episode-camera instances in the detailed benchmark section; these are different counts as stated in the supplied main-body extraction.
What the paper evaluates
The benchmark uses DROID, RoboChallenge, the AIST Bimanual corpus, RoboCOIN, and NVIDIA GR00T Teleop GR1. It evaluates DreamGen, DreamDojo, and Ctrl-World. The paper reports 291 single-view videos for embodied-physics human evaluation and 300 multi-view videos for multi-view consistency, graded on a three-point scale by two annotators with expert resolution. The evaluation includes Visual Integrity, PSNR, text alignment, and FVD, alongside KIVA and RIGIS. Speedup is measured as dense-baseline generation time divided by accelerated generation time per segment, excluding model-loading and VAE-decoding overhead. The paper reports 29,000-plus rollouts across nine acceleration methods.
Limit to keep in view
KIVA and RIGIS do not verify contact-level physical validity, which the authors say would require integrating full physics simulators.
arXiv:2610.31617
DiffDreamer: Coupling World Models with Differentiable Simulation for Agile Quadrotor Control
DiffDreamer combines a learned world model with differentiable quadrotor simulation so that external motion is predicted in latent space while the vehicle’s physical state is propagated with analytical gradients. The actor receives both implicit world-model features and explicit position, velocity, attitude, and angular-velocity states; its objective separates non-differentiable task rewards from differentiable motion rewards. On aggressive tracking and dynamic-gate racing, DiffDreamer converges faster and achieves stronger simulation results than DreamerV3, PPO, BPTT, MPC, and Elastic Tracker, with proof-of-concept transfer to a real quadrotor.
How it works
The controller is formulated as a POMDP with partial observations. DiffDreamer adopts a DreamerV3-style recurrent state-space model whose deterministic hidden state and stochastic latent feature encode history-dependent external dynamics. The simulator separately propagates the quadrotor’s explicit state using full nonlinear rigid-body equations, including position, velocity, quaternion attitude, and body angular velocity. At each step, one action is sampled from the actor and applied to both pathways.
What the paper evaluates
The paper evaluates two end-to-end benchmarks in VisFly, a high-fidelity PyTorch quadrotor simulator with explicit differentiable physics. Aggressive Tracking requires continuous tracking of a fast 3D target while keeping it centered in the camera field of view at a prescribed distance. Dynamic-gate Racing requires sequential traversal of moving gates in the shortest possible time. The policy outputs collective thrust and body-rate commands, and the simulator uses the full nonlinear rigid-body dynamics.
Limit to keep in view
Evaluation is limited to quadrotors; extending the approach to other embodiments requires suitable differentiable models and additional task-specific engineering.
arXiv:2610.36791
RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions
RynnWorld-Latent learns visual dynamics across human, robot, and simulation videos by conditioning a video diffusion transformer on continuous latent actions from a frozen RynnLAM encoder. Its global action tokens and frame-aligned local modulation let each transition descriptor influence the corresponding future video latent. After adaptation to a target robot, the model also accepts robot actions to support digital teleoperation and synthetic demonstration generation. The paper reports stronger prediction metrics on the evaluated public and real-robot benchmarks, and higher policy success rates in its reported simulation and real-robot settings.
How it works
The model predicts future video latents from the first latent frame and a sequence of latent actions, as formalized in the conditional prediction distribution. The training corpus spans 23 datasets and includes human egocentric video, robot demonstrations, and simulation rollouts. For each transition, frozen RynnLAM extracts a 608-dimensional descriptor with motion, global-action, and camera components.
What the paper evaluates
For public prediction benchmarks, the paper evaluates the DreamDojo GR-1 In-lab Eval, EgoDex Eval, and DreamDojo-HV Eval sets. Open-loop rollouts are conditioned on an initial frame and recorded native robot-action sequences; the last predicted frame autoregressively becomes the next conditioning frame. Higher PSNR and SSIM and lower LPIPS are better. RynnWorld-Latent reports the best listed values on all three sets, though baseline measurements are literature-reported and do not isolate the interface's contribution. For the authors' real-robot prediction benchmarks, the model is evaluated on held-out episodes from represented tasks on Marvin-WUJI and Astribot-S1.
Limit to keep in view
The quantitative public-benchmark baselines are literature-reported after robot post-training, so differences in backbone, data, and training prevent attributing the full gap to the latent-action interface.
Read the full reviews
- WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration · arXiv:2610.39721
- RoboWorM: Evaluating Movement Realism, Consistency and Acceleration Affects in Embodied Video World Models · arXiv:2610.33807
- DiffDreamer: Coupling World Models with Differentiable Simulation for Agile Quadrotor Control · arXiv:2610.31617
- RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions · arXiv:2610.36791
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.