Digest

Reset

1167 papers

2026-11-02 · arXiv:2610.42096

CausalWAM: Learning the Causes of Actions in World Action Models

Juntao Cheng, Jinzhou Tang, Fan Feng et al.

CausalWAM addresses a mismatch in world action models: predicting the whole scene is not the same as learning which latent causes drive actions. The authors prove that, with a sparsity constraint, action-relevant latent variables are identified up to permutation and the deployed compact state is an invertible function of the true action-relevant block. They propose CWAM by adding a sparse routing path to the action head of Cosmos 3. Experiments validate theory and improve closed-loop success efficiency.

DrivingAI summary
2026-11-02 · arXiv:2610.42146

Hierarchical World Models with Implicit Dynamics

Gaoyue Zhou, Yvonne Wu, Zichen Jeff Cui et al.

ID-WM (Implicit Dynamics World Models) is a hierarchical world model for long-horizon zero-shot planning. A high-level diffusion model generates neighboring reachable latent states using smooth Gaussian noise (implicit dynamics), while a deterministic low-level inverse dynamics model grounds those transitions into executable actions. The authors report 38.5% improvement over state-of-the-art policy and diffusion planners and 135% over single-level world model MPC, with videos at https://id-wm.github.io.

Latent DynamicsAI summary
2026-10-30 · arXiv:2606.15768

LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

Jialei Chen, Kai Wang, Kang Chen et al.

LaWAM proposes a latent-action-conditioned Latent World Model (LaWM) that predicts compact latent visual subgoals (future observation features) instead of generating future pixels/videos. The policy uses these predicted latent subgoals to produce dynamics-aware robot action chunks with low-latency, single-pass latent prediction. The authors report state-of-the-art or competitive success on LIBERO and RoboTwin and real-world manipulation.

RoboticsAI summary
2026-10-30 · arXiv:2610.37891

When More Planning Hurts: Learned Planners for World Models

Josef Chen

On PushT, a fixed supervised action updater (1,505 parameters) solves 8/16 reserved scenes. In the same reserved comparison, released LePlanner solves 16/16 with summed planning+encoding time 5.652s vs 59.079s for the updater. In a matched intervention, increasing early updates (4→8→12) yields 9/16, 8/16, 7/16 successes and extra updates do not rescue failures.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.37963

Self-Supervised Latent Action-Conditioned World Modeling for Time Series

Chia-Han Kuo, Chengshuai Zhao, Huan Liu et al.

ACTS-WM is a self-supervised latent action-conditioned world model for time-series. It infers discrete latent actions from observed time-series transitions without action annotations, using an inverse/forward dynamics setup with vector-quantized actions and task embeddings. It then pretrains a JEPA-based time-series encoder and an action-conditioned predictor to model transitions in latent space, evaluated on 4 tasks across 22 datasets.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.38266

Recursive Self-Improvement in Physical AI: Mechanisms, Evidence, and Evaluation

Xiaofeng Han, Kaize Luo, Zenghuang Fu et al.

The authors survey recursive self-improvement (RSI) in physical AI by separating learner updates from “improver” process revisions, and by evaluating when experience yields persistent capability gains, inherited process revisions, and verified improver gain. They propose common-data and interactive comparisons to isolate how experience is used versus acquired, and they connect evidence criteria to physical constraints like contact validity, recovery, retention, and resource costs.

Code availableAI summary
2026-10-30 · arXiv:2610.38289

PulmoWorld: An Asynchronous Multimodal Medical World Model for Longitudinal Pneumonia Dynamics

Cheng Tang, Xinyu Zhang, Jiawei Chen et al.

PulmoWorld is an asynchronous multimodal medical world model for longitudinal pneumonia dynamics. It uses history-aware filtering to propagate a shared patient belief forward in time and apply observation-driven corrections only when measurements are observed. The resulting belief supports recursive simulation for forecasts and evidence-constrained planning via Cap-MBRL, which searches within an authorized action set and uses risk-budgeted replacement of a frozen fallback.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.38439

Predict the Reward, Not the World: World Reward Models for ML Agents

Yuxiao Lee, Zherui Yang, Fan Liu et al.

The authors propose World Reward Models (WRMs) that learn to choose which candidate ML program will score higher using only real execution outcomes, without predicting intermediate execution traces. They prove limits of predicting execution and show how cross-entropy on executed winners can drive a WRM toward optimal selection. Their ForeCritic (from Kev-9B) improves pairwise selection accuracy and speeds up search-based agents while keeping scores similar.

RoboticsAI summary
2026-10-30 · arXiv:2610.38445

DES-JEPA: Learning a Shared Vocabulary for High-Dimensional Discrete Event Sequences

Hugo Math, Maximilian Forstenh\"ausler, Christos Anagnostopoulos

DES-JEPA learns a shared discrete codebook for discrete event sequences with unaligned, domain-specific event vocabularies by training on synthetic stochastic processes. DES-PFN reuses the frozen codebook for in-context next-event prediction (zero-shot, no task-specific fine-tuning). The authors report competitive zero-shot performance on eight real datasets and scaling to 700 distinct event types.

JEPAAI summary
2026-10-30 · arXiv:2610.38508

SaiVLA-JEPA: Evaluating Future Latent Prediction in JEPA-Conditioned VLA Control

Rufeng Xiao (VLA Research Team, Synthoid.ai), Ruihao Xiao (VLA Research Team, Synthoid.ai), Wei Chen (VLA Research Team, Synthoid.ai) et al.

SaiVLA-JEPA tests whether adding a JEPA-style future latent prediction objective improves closed-loop control for a vision-language-action policy that already uses frozen V-JEPA features. Two training-only residual heads predict future JEPA representations at horizons 4 and 16 and are removed at inference. On LIBERO-Plus, current JEPA conditioning yields higher success, while future supervision shows no detectable aggregate rollout gain at the evaluated training budget.

DrivingAI summary
2026-10-30 · arXiv:2610.38657

Explore Where It Matters: Targeted Exploration for Efficient Rover Adaptation

Jacob Levy, Arthur Bouton, Tristan D. Hassler et al.

TERRA is an online adaptation framework for planetary rovers that adapts a simulation-pretrained world model using a reward-weighted prediction loss and trains a loss predictor to target where future performance-relevant model errors are likely. It uses this signal to guide exploration during MPPI-based MPC. In simulation it reaches ~90% slope-climbing success within five trials under severe pretraining shifts, and on the ERNEST rover it recovers coordinated AGS locomotion within three to five real-world trials.

DrivingAI summary
2026-10-30 · arXiv:2610.38903

Structural-Test Geometry as an Inductive Bias for Partially Observable World Models

Todd Y. Zhou

The authors propose posterior-marginalized structural tests that induce a pseudometric over histories, then train quotient-preserving (QP) world models by aligning latent Euclidean distances to that metric. Across a fixed six-task diagnostic suite, QP improves planning success by 26.1 points over metric decoding and 6.4 over structured belief distillation under held-out structural variation.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.38980

Context Compaction for World Models

Hayder Tirmazi

The authors frame world-model consistency on revisits as a context retrieval problem, separating (1) model-layer context capacity (KV cache/attention) from (2) agent-harness context selection/compaction. They build a room-revisit benchmark and a context-compaction agent harness. They report that context compaction retrieves rooms up to 4× beyond the model context window without retraining, and that SWA can lose most context retrieval when dreaming.

2026-10-30 · arXiv:2610.39049

Reinforcement Learning for the Game of Balatro: A Fast Simulator and Two-Policy Training Approach

Ajay Vikram Periasami, Anjaneya Gupta, Ethan Cheng et al.

The authors present an end-to-end RL pipeline for Balatro using a Rust high-throughput simulator verified against the live client, dynamic action masking, and a two-policy multi-agent PPO approach: a tactical blind policy and a strategic shop policy trained jointly with per-agent reward routing and explicit cross-agent credit.

DrivingAI summary
2026-10-30 · arXiv:2610.39073

Certified Access: Test-Time Compute on Conservative Physics-Structured AI Memory

Pratik Jawahar, Riccardo Maggioni, Jonas Petersen et al.

The authors propose a framework for test-time compute on a Latent Dynamics Model used as memory, where latent access is implemented by structure-preserving symplectic dynamics. For two access mechanisms (a Lorentz squeeze and a gated canonical translation), they provide “certificate stacks” covering exact phase-volume Jacobians, bounded energy change, and latch-transport laws—validated on a designed analytic testbed. They also evaluate calibrated compute rationing and non-local routing limits against a physics-free baseline.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.39076

Tactus: Tactile & Pose Sensing Glove for Natural Dexterity Data Collection

Li Guan, Rachael Granberry, Sahar Aseeri et al.

Tactus is a textile-integrated sensing glove that combines calibrated fingertip tactile force sensing with IMU-fused hand pose tracking designed to remain accurate during hand-object occlusions. The authors provide a calibration pipeline (factory + per-session online) and introduce standardized tactile-force and pose benchmarking protocols, evaluated on Tactus and two commercial tactile gloves.

RoboticsAI summary
2026-10-30 · arXiv:2610.39100

Generating Critic-Guided Synthetic Experience in Latent World Models

Jorge de Freitas, Christos Ziakas, Alessandra Russo

The authors propose Critic-Guided Synthetic Experience (CGSE), which uses a frozen latent world model as a local transition generator for offline RL. Starting from latent states in an offline dataset, CGSE samples candidate actions, refines them to maximize a first-order critic approximation within a trust region, rolls out one step in the frozen world model, and appends the best synthetic transition to the replay buffer. No extra real-environment interaction or test-time planner is used.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.39232

Seeing and Imagining as One: Unified Video Generation and 4D Perception for Embodied World Modeling

Haitang Feng, Zhuoman Liu, Zhennan Wang et al.

GenSight is a unified embodied world model that combines future video generation and dense 4D perception (depth, embodiment segmentation, and first-frame-relative 2D flow) in a single pretrained video diffusion transformer. It uses a shared DiT backbone with lightweight task modulation, a common spatiotemporal (image-like) representation, and joint training in a unified velocity prediction space. In generation it produces future video with aligned dense predictions; in perception it predicts dense modalities in one step from an observed video.

2026-10-30 · arXiv:2610.39278

WorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining

Yuhao Lu, Jiongxu Chen, Hexiang Wei et al.

WorldFlow is a generative world model that jointly predicts future video and query-based 3D trajectories. It represents 3D motion as continuous displacement tokens aligned to the video grid and uses a unified Mixture-of-Transformers backbone with a rectified-flow objective to denoise both modalities. Trained on ~1.44M object-level samples, it achieves the lowest ADE/FDE among evaluated learning-based methods on WorldFlowBench, including SpatialCog (ADE 0.060 m).

2026-10-30 · arXiv:2610.39302

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Randall Balestriero, Yann LeCun

The authors prove a “Platonic Representation Theorem” for JEPAs: if modalities share the same positive-pair distribution formed from two observations of the same scene (“good views”), their learned representations match up to one rotation. They also show next-token prediction is a “bad-view” JEPA and validate the theory with scaling experiments from 145k captions to texts of 940M web images, reporting that good views are far more sample-efficient than next-token prediction and yield stronger cross-modal alignment.

JEPAAI summary