Digest

Reset

1167 papers

2026-10-06 · arXiv:2610.05349

Learning Conditional Source Distribution via Flow Reversal for Temporal Flow Matching

Kuan-Hsun Tu, Hsuan-Chi Liu, Jia-Wei Liao et al.

CNP-Flow is a flow matching (FM) framework for temporal generation that learns condition-dependent source distributions. It uses a conditional noise predictor (CNP) to output an isotropic Gaussian per temporal condition, supervised by flow reversal of a pretrained FM model. A three-stage pipeline pretrains FM, trains CNP via likelihood/KL losses, and fine-tunes FM with the learned source prior. The authors report improved generation quality on video prediction/interpolation and 7-DoF Franka motion planning.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.05359

Optimal Control with Learned Critics under Unmodeled State Dependencies

Philipp Schoch, Markus Ryll

The authors present a learning-based MPC that augments a nominal analytical model with a residual dynamics network to handle unmodeled state-dependent effects, and blends a learned action-value critic into a constrained iLQR planner to reduce finite-horizon myopia. A GPU-accelerated batched ALM-iLQR solver evaluates learned components inside the control loop and solves thousands of trajectory optimizations in parallel in simulation.

DrivingAI summary
2026-10-06 · arXiv:2610.05409

BeliefGraph-JEPA: Structured Latent World Models for Action-Conditioned Time Series

Yue Li, Kangqi Ni, Zhen Tan et al.

BeliefGraph-JEPA is a structured latent world model for action-conditioned multi-target time-series forecasting. It factorizes future-driver influence into typed latent-effect states that are rolled forward under future drivers, routed through a typed graph to target-specific nodes, and used as the JEPA-space predictive base. A capacity-controlled gated residual injects direct driver information. The authors report improved forecasting over known-future-covariate baselines on four multi-target systems.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.05418

EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation

Yuheng Na, Zhide Zhong, Junjie He et al.

EvoMem-VLA targets long-horizon, memory-dependent robot manipulation in VLA by adding state-evolution memory that explicitly encodes observed changes between historical states. It uses conditional delta tokenization for ordered frame-pair change tokens and a shared VLM backbone with task-adaptive routing (direct vs subtask route). The authors report success rates of 80.7% (RMBench), 82.0% (RoboMME), and 83.8% on real-world tasks.

RoboticsAI summary
2026-10-06 · arXiv:2610.05473

Hierarchical Reinforcement Learning with Stable Temporal Abstraction for Language Model Agents

Shayan Mohajer Hamidi, Yize Cheng, Yuanda Xu et al.

The authors study “temporal abstraction instability” in hierarchical language-model agents: the boundary policy can switch subgoals too often (premature replanning) or keep them too long (stale persistence). They propose STAC (Stable Temporal Abstraction via Constrained Optimization), which adds duration-dependent costs only to the sampled boundary decision, leaving other reward/advantage/critic signals unchanged. STAC improves success over a hierarchical baseline on ALFWorld and WebShop.

DrivingAI summary
2026-10-06 · arXiv:2610.05483

FLEX-WAM: Flexible Block-Causal World-Action Models for Long-Horizon Imagination and Planning

R. Khorrambakht, Joseph Amigo, F\'elix Lebel et al.

FLEX-WAM is a flexible, KV-cacheable block-causal world–action model for unified simulation and policy inference. It supports variable-length context, non-causal prediction horizons, and infinite autoregressive generation frame-by-frame or block-by-block. The authors use balanced state/action flow-matching gradients and FD elasticity to preserve action responsiveness. In MCTS imagination, it solves long-horizon PushT and all five OGBench Puzzle-4x4 tasks; on a bimanual robot, one checkpoint enables play and expected-outcome prediction.

2026-10-06 · arXiv:2610.05550

When Low Prediction Error Misleads Planning: Diagnosing Representation, Dynamics, and Decision Failures in Latent World Models

Rui Min, Xianyao Li, Fang Xu et al.

The authors show that the latent prediction-error component with the largest impact on MSE need not be the one whose correction improves action selection. Using a center/response split of endpoint error and matched-candidate tests across four tasks and four model families, they find MSE is dominated by center error in 14/16 cells, but in six cells correcting action-relative responses yields better physical rank correlation and top-30 elite quality.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.05711

From Pixels, Without Pre-training: Joint Generative and Self-Supervised Representation Learning in One Model

Vicente Balmaseda, Ching-Long Lin, Tianbao Yang

SCION trains a single pixel-space transformer to learn self-supervised representations (contrastive loss on clean augmentations) and to generate images via flow matching in one jointly optimized model. The shared encoder is conditioned on timestep and a conditioning embedding derived from either a learned null embedding (representation learning) or the image’s own [CLS] embedding (generation). A learned prior enables unconditional sampling without a reference image.

2026-10-06 · arXiv:2610.05731

T-JEPA: A Temporal Joint-Embedding Predictive Architecture for Learning Better Remote Sensing Representations

Bowen Peng, Li Liu, Yongxiang Liu et al.

T-JEPA learns remote-sensing representations from sparsely and nonuniformly sampled Earth observation sequences by predicting time-gap-conditioned latent transitions. A shared single-frame encoder produces masked source latents, and a temporal predictor estimates the complete target latent field conditioned on elapsed time, supervised across all temporally ordered pairs at multiple temporal scales. Asymmetric metadata injection plus masked pixel reconstruction improve learning and reduce shortcut learning.

JEPAAI summary
2026-10-06 · arXiv:2610.05739

HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models

Zhuokun Chen, Feng Chen, Xi Lin et al.

HLA-WM is a training-free hybrid linear-attention method for long-horizon video world models. It addresses long-range forgetting in Gated DeltaNet (GDN) by caching chunk-wise affine (Aᵢ,Bᵢ) transition summaries, retrieving scene-relevant chunks using camera geometry overlap, and recomposing them into a query-specific recurrent state. It improves long-range revisit consistency and camera control on SANA-WM-Bench and generalizes to MBench-A, with reduced memory and small throughput cost.

2026-10-06 · arXiv:2610.05861

Imagine to Act: High-Fidelity Data Synthesis via Image Editing World Model for Scalable GUI Agent Training

Yongxin Ning, Runliang Niu, Qianli Xing et al.

Infinite-Dreamer is a simulation-free GUI data synthesis method using a pixel-level Image Editing World Model. It treats GUI transitions as image editing guided by VLM-extracted structured delta-text, then fine-tunes an image editor to generate next screenshots and multi-step imaginary trajectories. The synthesized data trains Infinite-Actor (Qwen3-VL fine-tuned only on synthesized data), improving AndroidWorld and MobileWorld task success.

Video World ModelsCode availableAI summary
2026-10-06 · arXiv:2610.05912

MiniCorp: The Last Mile of the AI Agent Firm

Jingying Zeng, Zhenwei Dai, Jinning Li et al.

MiniCorp is a closed-loop office simulator where a persistent multi-role agent firm runs an e-commerce company against an evolving external market model. It records communications and decisions continuously, preserves information available at the time, and supports checkpointing to replay the same state under alternative decisions. The authors evaluate end-to-end external-world fidelity against empirical market-response patterns and show coordinated role behavior, adaptive learning from market feedback, and sustained advertising exploration with long-term strategic guidance.

DrivingAI summary
2026-10-06 · arXiv:2610.05969

Reachability-Aware Diffusion Policy Optimization

Hikmet Simsir, Kutay Demiray, Ozgur S. Oguz

RADPO (Reachability-Aware Diffusion Policy Optimization) is a model-free online diffusion-policy method for continuous-control safety. It learns a discounted first-hit reachability value to shape a reward critic, and uses an adaptive safety multiplier updated from realized episodic costs versus a budget. The diffusion actor is trained via value-weighted denoising regression on candidate actions, requiring no dynamics model, critic action gradients, or differentiation through the reverse diffusion sampler.

DrivingAI summary
2026-10-06 · arXiv:2610.05996

EpicWorldModel: Exploration-driven Planning with Latent World Models

Bowen Feng, Julian Ost, May Mei et al.

EpicWorldModel addresses ambiguity under partial observability in latent-space planning by training a stochastic JEPA world model. It uses a conditional flow-matching objective to generate multiple plausible future latent states, then estimates uncertainty via flow predictive variance (ensemble disagreement). This uncertainty is added to CEM planning to balance goal reaching with exploration; results show best or on-par performance and up to 22% higher success rate over LeWorldModel.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.06008

Ultrasound Operator Guidance Using World Modeling and Retrieval Based Action Planning

Noortje I. P. Schueler, Hans van Gorp, Ruud J. G. van Sloun

The authors propose retrieval-induced latent transition models for ultrasound operator guidance. Observations are encoded with a V-JEPA 2.1 backbone into a latent space, and guidance is generated by retrieving physically recorded transitions from a reference database and performing receding-horizon planning from the live ultrasound feed alone (no pose sensing at deployment). On carotid ultrasound, they report 86% target-view reachability in retrospective closed-loop episodes vs 52% and 43% baselines.

DrivingAI summary
2026-10-06 · arXiv:2610.06171

Controllable and Photorealistic Pedestrian Risky Motion Generation for End-to-End Driving Safety Evaluation

Siyuan Liu, Miao Li, Haibao Yu et al.

ControlPed generates photorealistic, motion-controllable vehicle–pedestrian safety-critical scenarios for end-to-end driving safety evaluation. It builds on HazardPed (from 10,352 traffic videos) to synthesize conflict trajectories, lift them into text-conditioned 3D human motions via diffusion, and render multi-view sensor observations using animatable 3D Gaussian Splatting avatars. Evaluating 88 rendered scenarios shows seven end-to-end driving models’ mean HDScore drops from 88.8 to 47.4.

DrivingAI summary
2026-10-06 · arXiv:2610.06235

Encoded but Not in Control: Revealing the Grounding Gap in Vision-Language Robot Policies

Shaohan Jiang, Jiahang Cao, Qiduo He et al.

The authors study scene-preserving instruction interventions where the instructed visible target changes but the scene stays fixed, for vision-language-action policies (VLA) and world-action models (WAMs). They report that even with instruction sensitivity and high linear decodability of the instructed target, evaluated policies predominantly approach and pick the original target named by the scene-associated prior, revealing a representation-to-control grounding gap.

RoboticsAI summary
2026-10-06 · arXiv:2610.06250

Generative World Models Enable Predictive Control of Laser Melt Pool Dynamics

Yiyang Yan, Markus Bambach, Mohamadreza Afrasiabi

The authors introduce a generative latent world model for localized 3D laser melt-pool dynamics, predicting evolution from histories of temperature and phase morphology under candidate laser actions. The model’s differentiable generative latent dynamics are used as a predictive control plant: MPC optimizes laser schedules via gradients through imagined futures, and an amortized latent policy produces actions in a single forward pass. They re-validate control schedules in high-fidelity CFD.

DrivingAI summary
2026-10-06 · arXiv:2610.06280

Future Anchored Verification and Online Recovery for World Action Models

Zhibin Qin, Zhenxiong Tan, Xinchao Wang

FAVOR (Future Anchored Verification and Online Recovery) uses a world action model’s own predicted future frames as anchors to (1) verify execution drift and (2) guide online recovery. An Anchor Verifier compares observations to their anchors (with actions and horizon context) to flag deviations. Anchor-Guided Recovery uses a vision-language model to issue a short corrective instruction to restore the anchored condition, then resumes the task.

DrivingAI summary
2026-10-06 · arXiv:2610.06349

KineWorld: Action-Induced Transport Fields for Embodied World Modeling

Ziying Song, Yuchen Liu, Zhuoran Xu et al.

KineWorld is an action-conditioned latent video world model for embodied prediction. It uses Kinematic Transport Lifting (KTL) to build renderer-derived, camera-aligned transport fields from commanded robot motion and injects them as clean structural conditions. Transport-Aware World Diffusion (TAWD) calibrates the transport support on the video-latent grid and reweights future-RGB flow-matching via a normalized mixture to emphasize action-responsive regions while preserving aggregate token weight. Trained on RoboTwin 2.0 ALOHA–AgileX data.

DrivingAI summary