Digest

Reset

1167 papers

2026-10-09 · arXiv:2610.12016

CausalDreamer: Learning Predictive World Models with Latent Disentanglement

Prince Jha, Nils Lukas, Kun Zhang et al.

CausalDreamer re-encodes a frozen Dreamer 4/MMBench2 video tokenizer latent into four groups factorized by controllability and reward relevance, routes actions only to the controllable groups, and predicts reward from only the reward-relevant groups. The pretrained dynamics model is fine-tuned to predict this factored representation. On MMBench2, it improves planning scores on clean tasks and their manipulated variants.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.12026

Humanoid World Action Model With Joint State--Action Generation

Yan Yang, Jikun Rong, Minzhao Zhu et al.

HWAM (Humanoid World Action Model) targets a humanoid action–execution gap by jointly generating high-level reference actions and the robot’s post-execution proprioceptive state. It trains with three conditional paths: Policy (state–action only), Forward Dynamics Modeling (future video conditioned on actions and realized states), and Inverse Dynamics Modeling (actions and realized states from visual transitions). The authors report the highest success rate among evaluated baselines on three LimX OLI tasks.

DrivingAI summary
2026-10-09 · arXiv:2610.12090

Recompose and Refine Latent Reasoning Flows for Vision-Language-Action Models

Hongyu Shi, Sen Zhao, Zuyu Zhang et al.

FlowMem is a unified vision-language-action model that stores successful latent reasoning as ordered “reasoning flow” fragments. At test time it dynamically retrieves and recomposes compatible latent suffixes across episodes, then refines the resulting reasoning tokens using current visual and proprioceptive evidence before conditioning continuous action generation.

DrivingAI summary
2026-10-09 · arXiv:2610.12099

UNITAS: A 3D-Native World Action Model for Embodied Manipulation

Ruixiang Wang, Yongyi Su, Wenlve Zhou et al.

UNITAS is a 3D-native world action model for embodied manipulation that unifies observations, actions, and scene dynamics in a shared metric 3D frame within each interaction. It represents action flow and scene flow as displacement trajectories in 3D, grounded by world-aligned 3D positional embeddings (with or without depth input) and encoded using a physical-time trajectory tokenizer. The authors report leading action-conditioned scene prediction and strong manipulation success, and release code.

RoboticsAI summary
2026-10-09 · arXiv:2610.12185

RESETTLE: Robotic Recovery through Disagreement-Triggered Retrieval and Efficient Corrective Control

Yuxin Chen, Senqiao Yang, Zixuan Wang et al.

RESETTLE is a model-agnostic recovery framework for frozen robot policies. It generates two action proposals from independent noise under identical conditioning and triggers recovery when their temporally weighted disagreement persists. It retrieves a same-task demo reference using an adapted V-JEPA encoder and executes one corrective action via a state-servo prior plus guarded visual residual, then returns to the base policy.

DrivingCode availableAI summary
2026-10-09 · arXiv:2610.12194

MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling

Jie Chen, Ruofei Bai, Yuxin Cai et al.

MiniWAM replaces dense native future-feature prediction in World Action Models with compact targets learned by PRISM (Predictive Representations via Inverse Spatiotemporal Modeling). PRISM compresses privileged current–future transitions using inverse-dynamics supervision plus feature reconstruction; its encoder is frozen, and MiniWAM learns to predict PRISM targets and actions from current observations. The authors report 65× fewer future tokens, up to 8× speedup in world-action training, and competitive results at 0.25B parameters.

DrivingAI summary
2026-10-09 · arXiv:2610.12231

Residual Modeling Closes the Regression and Generative Policy Gap in Robot Learning

Yuchen Zhou, Jiacheng You, Weikang Wan et al.

The authors analyze why action regression (MSE-Policies) underperforms generative policies in robot imitation: demonstration action residuals are input-dependent and heavy-tailed, and MSE gradients concentrate on large-residual samples. They propose HT-Policies, which use heteroscedastic Student-t residual modeling with learned input-dependent scales and single-pass inference. Across simulations and real robots, HT-Policies are competitive with generative baselines and faster in training/inference.

DrivingAI summary
2026-10-09 · arXiv:2610.12285

PLaW-VLA: Predictive Latent World Modeling for Vision-Language-Action Policies

Yu Liu, Hetian Guo, Tianlv Huang et al.

PLaW-VLA predicts task-relevant future latent states in a pretrained V-JEPA 2 representation space and conditions a VLA action model on these predicted futures using structured causal attention. Built on a Mixture-of-Transformers with three causally ordered experts, it improves long-horizon control and distribution-shift generalization (e.g., +11.8 pp on RoboTwin Hard Horizon III; +1.77 pp on zero-shot LIBERO-Plus) with parallel latent prediction and ~1/19 the inference latency of generative world–action modeling at comparable performance.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.12299

Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction

Dahyun Chung, Siyoon Jin, Hyunwook Choi et al.

ME-World is a multi-agent egocentric world model for synchronized first-person ego-stream generation with fine-grained embodied interaction. It jointly denoises multiple ego streams in one shared token sequence, uses shared action conditioning from all agents’ motions, and grounds generation with shared environment memory (stream-aligned geometric evidence plus cross-stream anchor references). The authors report improvements in shared-world consistency, action control, identity preservation, and video quality on real and synthetic benchmarks.

2026-10-09 · arXiv:2610.12333

RiCo: Neural Simulation of Rigid-Body Interactions via Local Contact Reasoning

Ruixiang Ouyang, Guanren Qiao, Fansen Meng et al.

RiCo is a neural simulator for rigid-body motion that reasons over sparse local contact neighborhoods rather than dense scene/object interactions. It represents each surface point’s contacts via nearby candidate points (from other objects and static environment) and uses intra-object point-transformer reasoning to predict rigid motion updates via anchors. On MOVi, the authors report up to ~35% lower position error and ~38% lower orientation error (plus reported contact-fidelity improvements) and zero-shot generalization to scenes with 270 objects.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.12341

Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition

Kaisen Yang, Qingle Liu, Kejin Wang et al.

The authors introduce AAArena, a benchmark for adversarial “agent-for-agent” improvement using adversarial heuristic learning (AHL): AI agents revise executable game policies and supporting software while keeping base model weights fixed. In 12 games with archived human programs, Opus5.5+Claude Code earns 6 gold medals (rank-1 in frozen human ladders). No evaluated configuration tops the remaining 6 ladders; performance is weaker for more complex rule specifications.

DrivingCode availableAI summary
2026-10-09 · arXiv:2610.12349

SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstruction

Ruijin Hua, Zichuan Liu, Zhuokai Zhao et al.

SplitJEPA is a reconstruction-free JEPA that learns an invariant/variant decomposition of a latent dynamical world directly in representation space. The authors prove that under stationary Gaussian predictive dynamics and a full-rank variation condition, SplitJEPA identifies invariant and variant subspaces (no observation decoder), up to independent within-block orthogonal transformations. Experiments on synthetic nonlinear systems and robotic manipulation support robustness and efficiency.

JEPAAI summary
2026-10-09 · arXiv:2610.12368

LiteNWM: Efficient Latent World Models for Onboard Visual Navigation in the Wild

Linkai Liu, Yuntian Zhang, Zhenshan Bing et al.

LiteNWM is a latent navigation world model that evaluates multiple candidate trajectories more efficiently by sharing visual encoding across candidates and jointly predicting action-conditioned future latent representations at multiple horizons (H1–H4). A learned proposal-aware scorer selects trajectories. The authors report 17.56% lower macro trajectory error vs NoMaD+NWM-XL offline, 128.00× end-to-end speedup on RTX 5090, and real-robot success improvement from 43.3% to 83.3% vs NoMaD.

DrivingAI summary
2026-10-09 · arXiv:2610.12374

AgentGarten: Code Worlds for Evolving Agents

Jiawei Chi, Shangchen Miao, Zhiyuan Shi et al.

AgentGarten is a framework for “code worlds” that combine simulator/game-engine programs with a shared real-time neural renderer. Engines maintain persistent state and execute explicit interaction rules, while the renderer generates visual observations from structured conditions (depth or normals) via a geometry-conditioned autoregressive video model distilled with Adversarial Forcing (distribution matching + exact replay + real-data adversarial supervision). Agents can perceive, act, and improve across rounds via distilled playbooks.

Video World ModelsCode availableAI summary
2026-10-09 · arXiv:2610.12407

Diffusion-Steering-Based MPC with a JEPA World Action Model

Shashank Hegde, Alexander Popov, Elie Aljalbout et al.

The authors propose LeWAM, a decoder-free JEPA world action model using a bidirectional transformer trained end-to-end in four modes (forward, backward, inverse, policy). For planning, they introduce diffusion steering (DS) for sampling-based MPC by searching in the policy’s rectified-flow noise space rather than action space. They report improved linear-probe alignment (ignoring visual distractors), matching acting performance to a matched flow-matching policy, and improved closed-loop MPC performance.

RoboticsAI summary
2026-10-09 · arXiv:2610.12412

WorldCast: Distributed Multiplayer World Models

Ziyang Ye, Junchao Huang, Evelyn Zhang et al.

WorldCast is a distributed multiplayer world model: each player runs a local client with a video generator and a state model, and clients exchange only shared player and scene state. A camera-aligned player state field guides where other players render, while a shared scene-state memory bank lets clients reuse observations. On Counter-Strike 2, the camera-aligned player field improves player rendering rate by over an order of magnitude, and scene state improves whole-round visual consistency.

2026-10-09 · arXiv:2610.12417

WOVEN: Weaving Visual World Modeling into Multimodal LLMs

Zheyu Fan, Yue Zhang, Mingkai Deng et al.

WOVEN introduces a training source and benchmark for visual transition reasoning in multimodal LLMs. It generates (s,a,s′) transitions using a video-pretrained generative model, labeling items by 20 scene types, 5 action types, and 8 reasoning types (36,076 items total). Evaluating 38 frontier MLLMs, the authors report a systematic deficit versus humans. Post-training MLLMs on WOVEN learn a transfer-capable shared capability and obtain gains on multiple external transition benchmarks.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.12435

VioLA: Learning Generalist Humanoid Control Policies from Human Data

Mert Albaba, Jens Bei{\ss}wenger, Anna Manasyan et al.

VioLA learns a generalist humanoid policy by predicting body and hand motion latents (instead of joint commands) in a shared latent space used by frozen human/robot motion encoders and decoders. On a Unitree G1, the authors report zero-shot locomotion success (100%) and zero-shot manipulation success (88.6%) without task-specific fine-tuning, using human demonstrations alone; code and checkpoints are released.

RoboticsAI summary
2026-10-09 · arXiv:2610.12459

WorldGuide: Goal-Directed Video World Model for Procedural Task Execution

Ankan Deria, Komal Kumar, Hisham Cholakkal et al.

WorldGuide formulates procedural video generation as closed-loop task execution in visual world space. Given an initial image and task goal, a ContextPlanner predicts the next atomic action or a DONE token from the generated visual state and recent clip-action history; a diffusion-transformer Executor renders the corresponding next video clip using hierarchical visual memory. The model is trained on step-level procedural demonstrations from WorldGuide Bench.

2026-10-09 · arXiv:2610.12461

OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs

You-Zhe Xie, Ting-Wei Chou, Yu-Hsuan Li et al.

OuroWorld converts any static 3D Gaussian Splatting (3DGS) scene into a 3D cinemagraph with vivid, diverse motion that loops seamlessly from any viewpoint. A vision-language model prompts a video generator to make an approximately looping reference video, which is lifted to multi-view videos. The method then fits an Inconsistency-Robust Periodic 4DGS using a Fourier-series deformation field (looping by construction) and a Grounded Drift Field (absorbs cross-view inconsistency).