Digest

Reset

1180 papers

2026-10-02 · arXiv:2610.00976

Variational Streaming Flow: Probabilistic Forecasting in Physical Time

Hans Hao-Hsun Hsu, Minseon Gwak, Soon Hoe Lim et al.

Variational Streaming Flow (VSF) extends Streaming Flow (SF) for probabilistic forecasting in physical time. SF uses a deterministic velocity field and outputs a single future from a fixed history. VSF adds a latent variable sampled once per rollout and kept fixed, producing multiple plausible futures while retaining SF’s physical-time efficiency. The authors report improved long-horizon predictive accuracy and distributional fidelity on deterministic and stochastic dynamical systems, and better goal-directed success when used as a JEPA-based world-model predictor.

Latent DynamicsCode availableAI summary
2026-10-02 · arXiv:2610.01001

Calibration-risk routing for controlled world-model adaptation

Yifan Zhang, Liang Zheng

The authors propose MC-WM for model-based RL under simulator-to-target shifts by fitting both a residual-corrected simulator ensemble and a direct target ensemble from disjoint fit/selection/calibration partitions. A calibration partition measures standardized one-step state-and-reward error to route the deployed family with lower calibration risk. A learned confidence signal plus deterministic validity predicates weight one-step imagined policy updates without rewriting physical rewards.

Latent DynamicsAI summary
2026-10-02 · arXiv:2610.01019

FutureWorlds: Learning Robotic World Models from Alternative Futures

Hao Wu, Shengju Qian, Weiyan Wang et al.

FutureWorlds learns robotic, action-conditioned world models by generating multiple candidate futures, maintaining candidate-specific histories, and learning from relative quality. It uses diverse beam search with bounded candidate memory and MemSPO, a GRPO-style update using group-relative video trajectory rewards. On RT-1, BridgeV2, and RoboCasa, the authors report LPIPS reductions for 32-frame predictions of 14.78%, 20.84%, and 9.12% versus the strongest baseline, with only 200 MemSPO updates further improving quality and enabling longer-horizon prediction.

RoboticsCode availableAI summary
2026-10-02 · arXiv:2610.01039

Bootstrapping Video Interaction Generation with Synthetic State Transitions

Jiho Jang, Jinyoung Kim, Nojun Kwak et al.

The authors introduce a pipeline to bootstrap video interaction generation using synthetic state transitions. They create 'start' and 'end' state images via a taxonomy-guided prompt and state-of-the-art image editing, then generate a coherent video using State-Guided Sampling (SGS) to reduce artifacts. They also propose an automated evaluation system aligned with human judgments. Fine-tuning a base video model on their dataset improves plausibility and interaction quality.

DrivingAI summary
2026-10-02 · arXiv:2610.01048

Network World Models as Environments for Algorithm Design on Complex Systems

Rishab Alagharu, Hongji Pu, Zeeshan Memon et al.

The authors propose an action-conditioned Network World Model that learns network diffusion dynamics under interventions over time, applies each intervention exactly, and predicts the outcome that follows. Combined with an LLM-based coding agent search loop that uses fast model rollouts plus rollout diagnostics/credit and counterfactual probes, they report matching or exceeding the strongest baselines in 138/141 settings and up to 14.5× faster rollouts than Monte Carlo simulation. Code will be released upon acceptance.

Latent DynamicsAI summary
2026-10-02 · arXiv:2610.01092

Ego2Act: Evaluating Goal-Directed Manipulation in Egocentric Video Generation

Patrick Amadeus Irawan, Iskandar Muda Rizky Parlambang, Rava Maulana et al.

Ego2Act evaluates whether video generation models can produce goal-directed egocentric manipulation videos with complete multi-step physical execution. The benchmark has 110 real-world tasks (2,640 videos) and introduces Ego2ActJudge, a reference-free, rubric-based evaluator that scores task completion and physics plausibility. The authors find models often skip dependent steps and fail fine-grained contact/persistent state modeling.

2026-10-02 · arXiv:2610.01162

PhysicsLENS: Diagnosing Physical Property Blindness in Video Generation Models

Isaiah Milkey, Som Sagar, Aditya Taparia et al.

PhysicsLENS is a robotics-grounded video generation dataset/benchmark that evaluates both physical plausibility and “physical-property grounding” for hidden properties. Using matched scenario pairs (same conditioning frame and task, varying the latent property in the prompt), the authors find that plausible-looking videos often ignore the stated property: 34 of 47 plausible unobservable videos do not adhere to the property, and stating the property lowers plausibility only slightly (0.28 points; not significant).

2026-10-02 · arXiv:2610.01171

Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation under Embodiment Mismatch

Yoshiki Takebayashi, Giovanni Perantoni, Hikaru Sasaki et al.

EF-GAIfO is experience-based feasibility-aware imitation from observation for state-only demonstrations when human motions may be infeasible for the robot. It trains a state-transition reconstruction model on the robot’s own experience to estimate feasibility via reconstruction error, then uses feasibility weights in the GAIfO discriminator objective. Feasibility estimates become more discriminative as robot experience accumulates. Validated in simulation locomotion and on a real Unitree Go2 quadruped object-reaching-and-picking task.

DrivingAI summary
2026-10-02 · arXiv:2610.01201

iSEE: Object Permanence Through Self-Supervision

Pramish Paudel, Ajad Chhatkuli, Luc Van Gool et al.

iSEE is a label-free, self-supervised object-centric video model for object permanence under occlusion. It adds temporal evidence normalization (TEN) to detect occlusion and hold a slot’s appearance, separates appearance and position into two streams, and trains a walker to follow hidden positions using only where the object reappears. On LA-CATER static, iSEE returns reappearing objects to their own slot after 86% of occlusions, and localizes hidden objects within 4.1 mAP of supervised RAM.

Latent DynamicsAI summary
2026-10-02 · arXiv:2610.01204

Autoregressive Drillhole Modelling Under Distribution Shift

Yihao Ding, Daniel Yitian Su, Yiran Zhang et al.

The authors introduce DrillBench, a public benchmark of 49,671 Western Australian drillholes for next-layer prediction and free-running autoregressive stratigraphic generation under a transfer spectrum (in-distribution, graded spatial shift, and cross-province). They find a transfer boundary: spatial/geochemical conditioning helps locally but degrades under stronger shift, while lithology-sequence autoregressive models transfer more robustly. They propose a backbone-agnostic recipe combining large-scale pretraining with spatial retrieval of neighboring lithology.

Code availableAI summary
2026-10-02 · arXiv:2610.01215

AutoGUIWorld: Image Generators as Visual World Models for GUI Agent

Cheng Yang, Yifan Wu, Yutao Huang et al.

AutoGUIWorld generates GUI interaction trajectories without deploying or running the underlying software. It samples initial GUI scenes from OS/visual/interface specifications, uses a planner to specify atomic actions and intended visual consequences, and iteratively edits screenshots with an image generator. It yields 79,266 grounded, filtered step-level samples across Ubuntu, Windows, macOS, and Chrome, and improves GUI-agent fine-tuning on OSWorld and ScienceBoard.

DrivingAI summary
2026-10-02 · arXiv:2610.01224

Supervise What Decides Success: Criterion-Aligned Auxiliary Losses for Latent World-Model Planning

Takumi Hara, Kanata Suzuki

The authors show that latent world-model planning can fail because the latent distance used for ranking does not reliably preserve the physical quantities that define success. They propose an auxiliary training loss that regresses the success-criterion quantities from both encoder and predictor latent outputs using a linear head (discarded at test time). This improves success rate on PushT and cube by +3.5% and +3.4% (absolute) and is statistically significant.

Latent DynamicsAI summary
2026-10-02 · arXiv:2610.01319

Cross-entropy optimization with prioritized constraints

Francisco Roldan Sanchez, Pau de las Heras Molins, David Fridovich-Keil et al.

The authors introduce TierCEM, a variant of the cross-entropy method that enforces strict constraint priorities directly in elite selection via sequential filtering of sampled candidates from highest- to lowest-priority constraint. When a tier would eliminate all candidates, TierCEM relaxes using the blocking constraint while preserving all higher-priority constraints. They evaluate on 2D navigation and contact-rich pushing, including learned world-model margins, showing that reversing priority order changes which constraints are violated under conflict.

DrivingAI summary
2026-10-02 · arXiv:2610.01373

Learning Commute-Time-Preserving World Models for Planning

Michael Hauri, Peter Buttaroni, Fabian A. Mikulasch et al.

Commute-Time-Preserving World Models (CTWMs) learn an action-conditioned latent world model for planning where Euclidean distances in latent space match commute-time distances in the environment. The authors combine residual (latent displacement) prediction with a log-determinant covariance regularizer to prevent collapse while preserving the required eigenvalue-dependent scaling. They prove recovery of the correctly scaled Laplacian representation under deterministic reversible dynamics and report improved goal-reaching performance vs LeWM using half the parameters.

Latent DynamicsAI summary
2026-10-02 · arXiv:2610.01559

Completion Aware Guidance for World Action Models

Seungyeon Kim, Junhoo Lee, Baekseung Kim et al.

World Action Models can generate visually plausible, action-consistent futures yet still omit the state transition needed to complete the task (“task-incomplete imagination”). The authors attribute this to an objective mismatch between fixed-horizon video generation and short-chunk control. They propose Completion Aware Guidance (CAG), a training-free sampling method that guides instruction conditioning toward completion, improving success and reducing task-incomplete imagination.

DrivingAI summary
2026-10-02 · arXiv:2610.01581

Evaluating Physical Consistency and Plausibility in Generative Scenario Models for Autonomous Driving

Manasa Mariam Mammen, Zafer Kayatas, Stefan Wagner

The authors introduce a five-layer evaluation protocol for generative scenario models in autonomous driving, targeting both internal representation interpretability and physical executability. Layers 1–4 assess latent/feature alignment, baseline advantages, decoder feature emergence, and latent controllability via intervention. Layer 5 checks generated trajectories against vehicle dynamics constraints using a Physical Validity Rate (PVR). They demonstrate it on a guided VAE and apply it to other generative models.

DrivingAI summary
2026-10-02 · arXiv:2610.01614

Oneira: From Open-Ended Generation to Open-World Interaction in Video World Models

Xindi Yang, Baolu Li, Liam Lee et al.

Oneira is an interactive video world model that maintains an explicit, extensible world state table updated by a coding agent. The agent grounds actions to object instances, plans interactions, updates object presence and interaction state (e.g., removal/recoloring), and extends the table with newly generated objects during exploration. The updated table is rendered into a coarse conditioning video for a pretrained video generator, enabling open-world interactivity and persistent state over long horizons.

2026-10-02 · arXiv:2610.01698

ActiveWAM: Evidence-Aware Active Vision for World-Action Models

Renjun Wu, Luzhou Ge, Xuesong Li

ActiveWAM treats active vision manipulation as an evidence-aware retain–acquire problem, unifying pan/tilt “head” control with bimanual manipulation in a world–action model. It uses training-time inversion to constrain a frozen video prior with task-bearing source evidence and visible temporal changes, enabling raw-input deployment without online inversion or candidate viewpoint ranking. It co-trains with future-video prediction and introduces RoboTwin-AV, a 50-task benchmark with executable pan/tilt control.

DrivingAI summary
2026-10-02 · arXiv:2610.01746

End-to-End Learning vs. Modular Architectures: Comparative Insights into Autonomous Driving Systems

Kartik B. Kapse

The authors compare end-to-end driving (raw sensory inputs to control commands) with modular driving pipelines (perception, cognition/decision, planning, control), summarizing trade-offs in interpretability, scalability, robustness, and real-world suitability. They further propose a Four-Dimensional Architecture Selection Framework (12 binary criteria) and validate it retrospectively on ten deployed autonomous driving systems, recommending the correct architecture for 7/10 cases (4 exact, 3 partial).

DrivingAI summary
2026-10-02 · arXiv:2610.01842

On the Divergence of Accuracy and Mechanism Consistency in Time Series World Models

Haochen Zhang, Jiaheng Guo, Zhen Xu et al.

The authors formalize time series world models (TSWMs) with state, actions, and exogenous inputs, and introduce mechanism consistency: whether changing a plan moves forecasts in the declared direction of known action→state mechanisms. On eight datasets with real actions, they find accuracy and mechanism consistency diverge; directional supervision improves mechanism consistency without changing MAE.

Latent DynamicsAI summary