Digest

Reset

1167 papers

2026-10-20 · arXiv:2610.24534

Structured Latent Dynamics for Long-Horizon Embodied World Models

Kevin Chen, Yu-Wei Zhan, Haoran Liang et al.

The authors propose HWM, a structured latent dynamics framework that encodes backbone latents into generalized coordinates and momenta, evolves them with dissipative Hamiltonian dynamics, and maps back via a learnable inverse interface. Multi-step transition supervision aligns structured rollouts with latent trajectories from real observations. Experiments show improved higher-order motion prediction, reduced long-horizon energy drift, and better downstream task success when integrated into existing world-model backbones.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.24571

Representations Have Trajectories: Characterizing Temporal Representation Dynamics in Neural Networks

Shivaprasad B. Gowda, Sparsh Mittal

The authors track representational geometry (stable and effective rank) throughout training and report three recurring rank trajectories: persistent high-rank, collapse-and-recovery, and persistent low-rank. They define Destructive Representation Collapse (DRC) as rank-one representations that fail to recover. In controlled synthetic settings, collapse-and-recovery is linked to successful optimization, while DRC is linked to failure; on real-world models the relationship is configuration dependent.

AI summary
2026-10-20 · arXiv:2610.24587

LIBERO-Pro2.0: A Benchmark for Evaluating the Robustness of Embodied Foundation Models

Xueyang Zhou, Yangming Xu, Qianjiang Li et al.

LIBERO-Pro2.0 is a robustness benchmark for embodied foundation models under distribution shifts. It adds 22 static perturbations and 20 runtime interventions across six dimensions, yielding 1,648 validated shifted task instances, plus 39,390 training trajectories from model-generated and human demonstrations with six rounds of human review. The authors evaluate 17 policies and report large robustness gaps under runtime shifts.

DrivingAI summary
2026-10-20 · arXiv:2610.24590

Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing

Yujia Yang, Quan Yuan, Jiawei Fan et al.

Scene-Adaptive VLA is an efficient VLA-based autonomous driving framework that uses dynamic layer routing to adjust active parameters based on real-time scenes. It characterizes model-perceived scene complexity using temporal scene variations and decision deviations encoded into scene-aware tokens, then uses a budget-conditioned routing mechanism to allocate an inference budget and select layer combinations. Adapted to SimLingo on Bench2Drive, it reduces inference-time computational overhead by 53.5% with negligible degradation.

DrivingAI summary
2026-10-20 · arXiv:2610.24619

Beyond Geometric Closure: Measuring and Diagnosing Mechanical Consistency in Visual Dynamics

Qibin Gu, Yunfeng Yan

The authors propose a six-criterion framework to measure mechanical consistency in visual dynamics, auditing geometric admissibility, reaction-free actuator effort, and work–energy balance in SI units. They evaluate KeyCLD and CON on 343 planar assemblies and V-JEPA+AC and LingBot-VA on 100 spatial assemblies, showing that lower image MSE does not imply native mechanical consistency, with calibrated replay revealing forecast errors far above measurement error.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.24662

From Prediction to Action: Population Health World Model for Digital Medicine

Xike Xie, Yu Li, Weizhong Yang et al.

The authors propose a Population Health World Model (PHWM) for digital medicine that represents an evolving health world where latent health states, population dynamics, interventions, constraints, and feedback are intrinsically coupled. They argue digital medicine should shift from static, prediction-focused models to action-oriented world models, requiring action-dependent representation learning, multi-scale intervention intelligence, and validation with explicit decision boundaries for bounded trustworthiness.

DrivingAI summary
2026-10-20 · arXiv:2610.24675

Action-Aligned Latent Planning for Versatile Humanoid Loco-Manipulation

Yutang Lin, Guanya Shi, C. Karen Liu

FIBER uses a frozen behavioral foundation model (BFM) latent space as a shared executable representation for diverse humanoid loco-manipulation skills. It projects each independently acquired interaction policy into BFM latents via action-aligned latent policy projection (LPP), then trains a single perceptive flow planner offline from aggregated closed-loop latent rollouts. An adaptive-horizon BFM (AH-BFM) supports both precise control and longer-horizon behavior.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.24681

MM-LeJEPA: Correlated Gaussian Regularization for Multimodal Joint-Embedding Predictive Architectures

Haowen Bai, Rui Gong$, Meichen Liu et al.

MM-LeJEPA extends LeJEPA to multimodal (audio-visual) joint-embedding predictive architectures by replacing LeJEPA’s isotropic task prior with a multimodal task prior that models independent information and shared content across modalities. Under stated assumptions, it derives a correlated Gaussian target with a single cross-modal correlation hyperparameter ρ, enforced via SIGReg using a fixed whitening transform. Code will be released.

JEPAAI summary
2026-10-20 · arXiv:2610.24914

Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design

Boyi Liu, Yite Wang, Xinyu Lian et al.

Arctic-SWE is a fully asynchronous RL framework built via model–system co-design, using long training on a small overfit task pool to expose instabilities and guide fixes. No single component is new, but together they form a reproducible baseline. On unchanged R2E-Gym tasks, Arctic-SWE trains Qwen3.5-4B to 62.0% pass@1 on SWE-bench Verified, without new data algorithms.

DrivingAI summary
2026-10-20 · arXiv:2610.24981

Dream What I Can: Sparse Visual Planning for Long-Horizon Robot Manipulation

Chongkai Gao, Wenlong Huang, Haonan Chen et al.

DreamCan is a hierarchical policy for long-horizon robot manipulation that plans with sparse visual subgoals grounded by reachability and task utility. From action-free robot video datasets, it learns dataset-supported reachability and task progress, uses dynamic programming to select subgoal sequences without manual annotations, fine-tunes a pretrained video model to generate subgoals, and executes them with an independently trained goal-conditioned low-level policy.

RoboticsAI summary
2026-10-20 · arXiv:2610.25013

Latent World Models Should Plan with Distances Along the Latent Manifold

Zunzhi You, Linwei Tao, Siyu Xu et al.

The authors study LeWM-style latent goal planning that uses Euclidean (L2) distance between predicted terminal and goal latents. They find L2 is reliable only near the goal and forms spurious low-cost basins farther away. They hypothesize latents lie on a curved manifold and propose Neighborhood Graph Cost (NGC) using shortest-path distances on a mutual nearest-neighbor graph over training latents; this changes only the terminal cost (no retraining) and boosts success rates across tasks.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.25163

M-JEPA: Predictive Self-Supervised Learning for Molecular Graphs with Scaffold-Shift Evaluation on Tox21

Karthik Iyer, Nasser Sabar

M-JEPA adapts Joint Embedding Predictive Architectures to molecular graphs by masking chemically coherent connected subgraphs and training a context encoder to regress embeddings of an EMA teacher. The authors introduce a leakage-resistant, compute-matched three-phase evaluation and test on scaffold-shifted Tox21, reporting discrimination (ROC-AUC/PR-AUC) plus calibration (Brier/ECE) and motif-level integrated-gradient attribution.

JEPAAI summary
2026-10-20 · arXiv:2610.25170

World Models, Trees and Foundation Models for Retail Forecasting: Evidence from Four Public Data Sets

Karthik Iyer, Madan Dabbeeru

The authors compare a plan-aware world model (roll a carried latent item state forward under known price/promotion plan) against the same network without roll-forward, LightGBM, Temporal Fusion Transformer, three pretrained forecasters (TimesFM, Chronos-Bolt, Timestral), seasonal naive, and an equal-weight ensemble. Best model depends on the product-store-time grain: the average is best on item-by-store and item-across-stores, while cost/accuracy patterns differ at coarser grains.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.25245

GeoV2V: Geometry-Grounded Video Diffusion Model for Driving Scene Generation

Yuchen Xi, Yuhan Guo, Chenwei Hou et al.

GeoV2V is a geometry-grounded video-to-video diffusion model for driving scene generation under novel trajectories (e.g., lane shifts). It integrates LiDAR- and depth-derived geometric priors with full-source video conditioning in the Wan 2.1 backbone. The authors introduce Para4D, a synthetic multi-lane synchronized dataset, to explicitly supervise lane-shift generation and improve cross-lane geometric consistency.

DrivingAI summary
2026-10-20 · arXiv:2610.25289

The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability?

Ronen Schaffer, Hila Manor, Tomer Michaeli

The authors study patch-token (token-decoding) explainability in Vision Transformers and find it is not guaranteed. Across many real-world ViTs, some models show aligned, others inverted, and some unexplainable maps, and the behavior can change with the optimization landscape (including task/data/optimization choices). They also analyze contrastive ViTs/VLMs and provide theory showing a toy transformer can implement the same optimal rule with different patch-decoding interpretability patterns.

AI summary
2026-10-20 · arXiv:2610.25296

World Model Interpretability: From Decoding Representations to Understanding Latent Worlds

Jize Wang, Yongcheng Jing, Xinyi Le et al.

The authors argue that JEPA latent spaces should be interpreted as latent world models, not just decoded representations. They propose a four-stage agenda—Ground, Structure, Verify, Act—to move from (1) mapping latents to world variables, to (2) recovering relations/dynamics, to (3) causally verifying predictive mechanisms via interventions, and (4) using verified understanding to diagnose and improve model behavior.

JEPACode availableAI summary
2026-10-20 · arXiv:2610.25333

CoRe-WM: Conditioned Response-State World Modeling for Longitudinal Treatment-Response Prediction in Breast DCE-MRI

Lin Zhao, Bowen Chen, Qiaohui Gao et al.

CoRe-WM formulates longitudinal breast DCE-MRI pCR prediction as outcome-free conditioned next-latent prediction. It learns a patient- and treatment-conditioned latent state-transition model using a JEPA-style objective, plus a factorized response state pathway supervised by image-derived response guidance. After pretraining, parameters are frozen and a linear readout predicts pCR at each decision point. Evaluated on I-SPY2 (5-fold) and transferred to I-SPY1.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.25360

Operator-Mismatch Contrastive Learning for World-Relative Trajectory Coherence

One Octadion, Firmansyah Sundana, Novanto Yudistira

The authors formalize “operator-blindness” in world-relative trajectory scoring: a scorer can reject corrupted futures and still order current-world vs other-world valid futures at chance, and standard generic top-1 evaluation can hide this. They propose Operator-Mismatch Contrastive (OMC) learning using valid-but-wrong-operator rollouts as semantic hard negatives. On synthetic IID/OOD and CoPhy feature-level validation, OMC improves operator-sensitive coherence scoring while keeping generic top-1 near 0.99.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.25414

UniState: Unifying Memory and Intervention for Controllable World Simulation

Jaeseok Jeong, Kyungmook Choi, Sohyun Chung et al.

UniState is a controllable world simulation framework that unifies spatial memory and user interventions in a shared world-state representation. Users specify object motion via 6-DoF transformations in 3D memory, while independently choosing the camera trajectory for observation. The video model follows the specified object motion and camera path using training/sampling (including Scene-Anchored Joint Guidance) that reduce ambiguity between object and viewpoint effects.

DrivingAI summary
2026-10-20 · arXiv:2610.25511

Dynamics-Aligned Representation Learning under Induced Partial Observability

SM Mazharul Islam, Manfred Huber

The authors propose stochastic spatial gating for latent dynamics models in model-based RL. During training, they mask pixels using a binary mask sampled from a Beta–Bernoulli random field (mask probability resampled each timestep) to induce controlled partial observability. They minimally integrate the method into RSSM by concatenating the gated observation and the mask to the encoder and adapting the reconstruction loss to only surviving pixels. Evaluation uses MPC-CEM open-loop rollouts on DeepMind Control Suite tasks.

Latent DynamicsAI summary