Digest

Reset

1167 papers

2026-10-14 · arXiv:2610.17201

Multics: Generative Multi-Player World Model with Hierarchical Communication

Kaiyi Huang, Wangbo Yu, Yukun Huang et al.

Multics is a multi-player video generation model that uses hierarchical communication (player, group, global) to improve cross-player consistency and controllability while limiting interference between unrelated local events. It introduces shared group tokens for intermediate aggregation and uses budget-constrained dynamic grouping based on recent camera trajectories. The authors report state-of-the-art results on 4- and 6-player settings and generalization from 4-player training to 6-, 8-, and 10-player inference without retraining.

2026-10-14 · arXiv:2610.17434

World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models

Junyi Wu, Zekun Li, Chenhongyi Yang et al.

World-T3 is a camera-controlled video world model that uses test-time training (TTT) as persistent, pose-indexed memory. During long-horizon generation, evicted blocks are consolidated into fast-weight parameters via a recursive least-squares rule, and later reads retrieve prior scene features by camera pose. The autoregressive model is distilled into a 4-step generator for real-time interactive streaming.

2026-10-14 · arXiv:2610.17526

Box-WAM: Box Trajectories as Compact State Representations for World Action Models

Qinxi Yu, Wenzhou Ding, Gloria X. Wang et al.

BoxWAM replaces dense future video prediction in world action models with oriented 3D bounding box (OBB) trajectories of task-relevant objects. The model jointly denoises future box tokens and action tokens in a rectified-flow transformer with causal state-action attention. The authors report lower future-modeling cost and improved action prediction, evaluated in simulation (RoboTwin2.0) and on a real robot.

DrivingAI summary
2026-10-14 · arXiv:2610.17584

RAN-JEPA: An Action-Conditioned World Model for Energy-efficient Mobile Network Control

Binghan Wu, Kai Zeng, Kai Yang et al.

RAN-JEPA is an action-conditioned world model for energy-efficient mobile network (RAN) control that addresses an observation–intervention asymmetry. It first learns a predictive latent representation from observation-only telemetry (and adapts it to a target deployment), freezes the encoder, then learns action-conditioned latent dynamics from controlled sleep/wake trajectories. Experiments use public operator data plus a real 4-cell 5G testbed.

DrivingAI summary
2026-10-14 · arXiv:2610.17593

Training Generalist Navigation Policies via Action Self-Distillation

Bowen Feng, Keqin Wang, Yulong Yang et al.

Action Self-Distillation (ASD) trains a language-conditioned navigation policy from trajectory-only robot datasets by using an open-source VLM as both a self-teacher and student. The self-teacher uses privileged ground-truth future trajectories to infer a language goal and structured reasoning; the student learns from onboard observations alone. Across 13 real-world indoor/outdoor tasks, the authors report improved success vs the strongest baseline.

DrivingAI summary
2026-10-14 · arXiv:2610.17625

4LC: 4D Latent Cache for Efficient Conditioning in Dynamic World Modeling

Jianhui Zhang, Zuyao Lin, Pengxin Xu et al.

4LC introduces a 4D latent cache that reuses static and dynamic visual representations for dynamic video world modeling. It stores composite features from multiple visual encoders bound to shared geometry and motion, and provides a spatiotemporal interface to read cached conditions at independently specified times and viewpoints. 4LC-Video uses these cache readouts for multiple video generation tasks using a single checkpoint without task-specific fine-tuning.

2026-10-14 · arXiv:2610.17630

From Pixels to Executable Physics: Behavior-Grounded World Models for Embodied Action Consequence Prediction

Ruo Chen, Jiahua Pang, Yingjie Niu et al.

DR-WM is a physically consistent world model for quadruped action consequence prediction. From an initial RGB frame, it reconstructs geometry and predicts dense, normalized priors over terrain contact properties (friction, roughness, stiffness). Calibrated with real action–response data, it simulates robot motion with spatially varying contact, derives camera trajectories, and guides video generation. Generated observations are used to iteratively update the terrain for long-horizon rollouts and VLM action selection.

RoboticsAI summary
2026-10-14 · arXiv:2610.17638

Acting Through Registers

Yiwei Zhang, Yantai Yang, Jin Gao et al.

World Action Models (WAMs) can show persistent high-norm spatial outlier tokens whose action relevance persists even after K/V projection attenuates their magnitude. The authors propose RegisterWAM: add learnable non-spatial register tokens for action conditioning, and pair them with observation flow-matching (FM) supervision to suppress spatial outliers while concentrating high-norm responses in registers. They also add noised-observation FM for recovery. RegisterWAM reaches 81.6% on LIBERO-Plus and 93.0% on RoboTwin 2.0.

DrivingCode availableAI summary
2026-10-14 · arXiv:2610.17651

Code2Worlds-2: Physically Plausible Agent Harness for Embodied World Simulation

Haoyu Zhang, Siming Li, Zeyu Zhang et al.

Code2Worlds-2 extends Code2Worlds to produce executable, simulator-compatible embodied 4D worlds. It appends (1) a SimReady Scene Compiler that adds collision geometry, estimates physical attributes, infers a support graph, applies contact/dynamics constraints, and gravity-settles before exporting MuJoCo MJCF, and (2) a Simulator-in-the-Loop refiner that converts visual dynamics into executable rigid/articulated dynamics and refines using simulator critics. It introduces the SimCode4D benchmark.

DrivingCode availableAI summary
2026-10-14 · arXiv:2610.17688

OccWAM: Occupancy World-Action Models with Joint Scene-to-Ego Flow Matching

Qihang Huang, Wenhui Huang, Collister Chua et al.

OccWAM is an occupancy-native world–action model that jointly generates dense metric future occupancy and continuous ego trajectories using scene-to-ego two-stage conditional flow matching. Stage I predicts scene-centric occupancy plus an absolute ego trajectory in a fixed current-ego-anchored frame; Stage II re-indexes the predicted world into future ego frames and jointly refines ego-centric occupancy and motion for action-aligned planning. It avoids pixel synthesis and preserves semantic geometry.

DrivingAI summary
2026-10-14 · arXiv:2610.17726

TSRL: Trigger-based Supervised Reinforcement Learning for Continual Agent Training

Yuqian Fu, Kaiwen Jiang, Zijun Song et al.

TSRL (Trigger-based Supervised Reinforcement Learning) uses RL as the optimization backbone and activates expert SFT only on “hard states” where the policy fails, detected via Group Success Rate (GSR) below a threshold. The authors theoretically bound forgetting by restricting supervision to failure states, and report consistent gains across ALFWorld, WebShop, and Deep Search at 7B and 32B scales, with up to 9.6 points more prior capability retention than SFT→RL baselines.

DrivingAI summary
2026-10-14 · arXiv:2610.17843

Physics-Aware Object-Centric SE(3) Forecasting for Real-Time Proactive Dynamic Grasping

Yirui Zhou, Hongli Xu, Ying Zheng et al.

The authors propose PAF-Grasp, an object-centric forecast-then-select framework for real-time proactive dynamic grasping. Instead of dense video prediction, it forecasts contact-relevant rigid-body trajectories in SE(3) from short pose histories, then selects reachable contact states under kinematic and timing constraints. Trained in randomized physics simulation, it transfers to real objects after recovering geometry and inertia. They report higher success in both simulation and real-world tasks with 15 ms forecasting.

DrivingAI summary
2026-10-14 · arXiv:2610.17903

DriveWLA: Unifying Future Dynamics and Scene Semantics through World-Language-Action Modeling for Autonomous Driving

Shuzheng Wang, Haoxuan Li, Shengzhen Li et al.

DriveWLA is a unified World-Language-Action model for autonomous driving that uses three MoT “experts” (World for future prediction, Language for semantic understanding, Action for trajectory generation) with Cross-Expert Attention to control information flow and align representations. The Action Expert is refined with Flow-GRPO reinforcement learning post-training beyond imitation learning. The authors report strong open/closed-loop performance and zero-shot generalization across driving simulators and datasets.

DrivingAI summary
2026-10-14 · arXiv:2610.17922

WoVR+: World Model as Reliable Simulator for Versatile VLA Policy Optimization with RL

Shangqing Zhou, Zhennan Jiang, Ning Ma et al.

WoVR+ extends WoVR’s hallucination-aware WM-RL to post-train Vision–Language–Action (VLA) policies with richer multi-view inputs and long-horizon, multi-stage tasks. It uses a VLA-interface-aligned multi-view world model with geometry-grounded action/camera conditioning and command-based proprioceptive updates, plus chunk-wise Flow-SDE exploration. For long horizons, it uses stage-local imagined training with stage-entry resets, stage-specific rewards, local horizons, and controllable sampling, improving success over SFT on RoboTwin and 6 real-robot tasks.

DrivingAI summary
2026-10-14 · arXiv:2610.18038

Data-Free Continual Learning for World Action Model with Rectified Generative Replay

Kaipeng Fang, Hangxu Liu, Chaojian Shi et al.

The authors propose R^2-WAM, a data-free continual learning method for world action models. A WAM regenerates rollouts of past tasks from their instructions, but raw rollouts cause forgetting due to missing task-completion signals and observation–action drift. R^2-WAM truncates rollouts at task completion and filters rollouts using physical-consistency flow scores, reducing forgetting by over 90% vs sequential fine-tuning and nearly matching ER without prior-task data.

RoboticsAI summary
2026-10-14 · arXiv:2610.18082

Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents

Xiyang Wu, Zhichao Liu, Shengxin zhang et al.

BeliefWeaver is a long-horizon failure-recovery framework that keeps a persistent relational belief state over repair frontiers across rollback, and uses an implicit world model (IWM) to evaluate repair hypotheses and prospective progress. Execution is gated by diagnostic evidence and ranked by belief. The framework is distilled into a 9B student via SFT and on-policy distillation (OPD) to reduce teacher inference reliance.

Latent DynamicsAI summary
2026-10-14 · arXiv:2610.18093

LynxScenes: A Large-Scale Real-World Benchmark for Outdoor Wheel-Legged Robot Perception, Navigation, Imagination and Reconstruction

Jun Tian, Hao Wu, Ruo Chen et al.

The authors introduce LynxScenes, a large-scale real-world dataset and benchmark collected with a LYNX M20 wheel-legged robot for outdoor, socially compliant navigation and scene understanding. It provides ~100K non-overlapping 30-s clips (833.6 h, 3,294.7 km, 3.0M annotated frames) with synchronized multimodal sensing and dense annotations, supporting instruction-conditioned trajectory prediction, BEV perception, future scene prediction, and multi-view reconstruction.

DrivingAI summary
2026-10-14 · arXiv:2610.18135

SkipWorld: Latent Skip Planning with Distilled World Models for Efficient Visual Navigation

Yifei Yang, Zehua Fan, Zike Yan et al.

SkipWorld is a latent skip planning framework for efficient visual navigation. It trains Latent AutoRoll for closed-loop latent rollout, distills it into SkipDistill for long-stride endpoint prediction, and uses Skip-then-Refine Planning to coarse-filter many candidates with the skip model and rerank selected ones with the autoregressive teacher. On RECON, HuRoN, SCAND, and TartanDrive, it reports lower average ATE/RPE and faster planning than NWM.

DrivingAI summary
2026-10-14 · arXiv:2610.18167

Beyond Planning: Differentiable Control-in-the-loop End-to-End Autonomous Driving

Zhidong Wang, Zirui Li, Zitong Shan et al.

DIVA (Differentiable Identified Vehicle dynamics for end-to-end Autonomous driving) embeds an identified actuator-level vehicle dynamics model into an end-to-end autonomous driving training loop, enabling a differentiable chain from perception to vehicle states. It replaces the external controller with a learned control decoder supervised on rolled-out vehicle states, uses rate-consistent control-input knots (3.8 ms control-only cycle), and applies on-policy re-identification plus an online disturbance observer for mismatch robustness.

DrivingAI summary
2026-10-14 · arXiv:2610.18174

Looped Reasoning Finishes Earlier Than You Think

Sumin Park, T. Konstantin Rusch

The authors show that slowly decaying oscillatory modes can become rate-limiting for convergence in recurrent reasoners by transferring their convergence rate to coupled complementary dynamics. Instead of damping oscillations, they propose QRR (Quotient Recurrent Reasoner), a gauge-equivariant recurrent architecture with exact token-local phase-rotation symmetry that leaves quotient (reasoning-relevant) dynamics unchanged by symmetry motion. Empirically, QRR converges stably and faster and matches or surpasses recurrent reasoning baselines on Sudoku, Maze, and Mini-ARC.

DrivingAI summary