Digest

Reset

1167 papers

2026-10-06 · arXiv:2610.04494

DreamTest: World-Model Surrogates for Search-Based Testing of Deep Reinforcement Learning Agents

Qinghua Xu, Guancheng Wang, Boxi Yu et al.

DreamTest is a world-model surrogate for search-based testing of deep RL agents. It adapts a recurrent state-space model (RSSM) to learn agent behaviour and environment dynamics from per-step training logs, then scores candidate environment configurations by imagined rollouts and a termination-weighted failure score. The authors evaluate on Parking, Humanoid, and DonkeyCar for failure prediction, test generation, and failure diversity.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.04539

Action-Consequence Alignment for Reliable Planning and Self-Improving in Latent World Models

Jinping Wang1, Zhiqiang Gao, Xiantong Zhen et al.

The authors show that latent action-conditioned world models can assign lower prediction error to nearby counterfactual actions that lead to physical outcomes farther from a recorded goal, creating a prediction–planning gap. They propose Action–Consequence Alignment (ACA), which uses a local action search and penalizes the searched actions’ prediction-error advantage over factual actions. ACA improves planning performance and reduces real goal error, and can guide additional data collection for self-improvement.

Latent DynamicsCode availableAI summary
2026-10-06 · arXiv:2610.04540

DreamFormer: Dream Imitation with a Transformer World Model for Language-Conditioned Robotic Manipulation

Mostafa Kotb, Cornelius Weber, Muhammad Burhan Hafez et al.

DreamFormer is a model-based, language-conditioned robotic manipulation agent that learns a task-agnostic Transformer world model from unstructured play data, then trains a policy inside the world model’s latent imagination using Dream Imitation with an intrinsic reward aligning imagined rollouts to expert demonstrations. It encodes full-resolution multi-view observations into a single token per timestep via a scaled categorical VAE. On CALVIN, it outperforms LUMOS and nearly doubles HULC on zero-shot transfer.

DrivingAI summary
2026-10-06 · arXiv:2610.04585

Frozen in a Frame: The Velocity Blind Spot in JEPA World Models

Tinghe Zhang, Chunyu Liu, Yu Leon Liu et al.

The authors argue JEPA world models trained from single rendered frames have a structural “velocity blind spot”: without motion blur, a still frame encodes configuration but not instantaneous velocity. They verify on official LeWM checkpoints that linear velocity probes are at chance while position probes reach R^2≈0.94–0.98. They propose RateIdent and TI-JEPA, splitting latent into pose and explicit finite-difference motion, improving stop-at-goal planning and fixing velocity-related diagnostics without privileged supervision.

JEPAAI summary
2026-10-06 · arXiv:2610.04607

ForeAct3D: Policy-Grounded Future World Modeling for VLA Policies

Zhe Tao, Feiran Wang, Gaowen Liu et al.

ForeAct3D is an action-conditioned semantic 3D future world modeling framework integrated into VLA policies. It decodes depth, 3-class semantics, and camera pose from policy features at current and future endpoints, grounds future queries on the policy-generated action chunk, and enforces physical consistency via background staticity, instance rigidity, and wrist-camera pose anchored to end-effector kinematics. No future prediction is required at inference.

RoboticsCode availableAI summary
2026-10-06 · arXiv:2610.04681

RoboIRS: Inference-Time Internal Representation Steering for Generalist Robot Policies

Jiuzhou Lei, Chang Liu, Dayou Li et al.

RoboIRS is an inference-time internal representation steering method for frozen robot policies. It uses successful vs. failed rollouts to train linear classifiers, selects outcome-relevant intervention sites, and injects classifier-derived steering directions into transformer residual streams without updating policy parameters. On 15 LIBERO-PRO simulation tasks with a frozen π0.5 policy, RoboIRS raises average success rate from 44.4% to 66.2%. It also improves real-robot manipulation and transfers to Cosmos Policy.

RoboticsAI summary
2026-10-06 · arXiv:2610.04767

Flow Policies as Actions of Skill-Level World Models: Learned and Symbolic Abstractions for Long-Horizon Planning

Andreu Matoses Gimenez, Andrei-Carlo Papuc, Chris Pek et al.

The authors build a skill-level latent world-model framework where one world-model transition corresponds to one execution of a flow-matching policy over complete skills. They compare four skill action abstractions—compressed seed, learned code, object-centric learned code, and a symbolic label—for long-horizon simulated block rearrangement (up to 14 sequential skills). The symbolic label achieves over 90% success up to six skills; without labels, the object-centric code matches it on single-skill tasks and retains half to three quarters of its success for two to five skills.

DrivingAI summary
2026-10-06 · arXiv:2610.04778

Repeated-Measure Leakage, Distribution Shift, and Reliability under Partial Observation in Patient World Models

Arjun Subramanian

The authors test leakage-resistant patient generalization and reliability for short-horizon gait digital-biomarker forecasting on GaitPDB. They build an evaluation ladder that progressively removes raw temporal overlap, same-recording familiarity, and same-patient familiarity before unseen-patient testing. GRU NMSE increases from 0.1227 (random-window) to 0.1393 (no raw overlap) to 0.1961 (patient holdout), and missingness degrades performance while MC-dropout variance decreases.

DrivingAI summary
2026-10-06 · arXiv:2610.04851

CURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery

Tao Feng, Fangxu Yu, Zijie Lei et al.

The authors propose CURIO, a curiosity-driven test-time learning framework for open-ended discovery. CURIO trains an Intrinsic Curiosity World Model (ICWM) to model token-level transitions in the policy’s hidden-state representation and adds a prediction-error bonus at sampled tokens outside the policy’s top-k choices. Epoch normalization and an annealed curiosity weight regulate the bonus contribution during test-time RL.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.04896

Tackling Sim-to-Real Mismatch Through Sampling-Based Disturbance Observers: From Analytical Models to Learned World Models

Tianqi Zhu, Jun Yang, Jianliang Mao et al.

The authors introduce SDOB (sampling-based disturbance observer) to extend disturbance-observer compensation to modern predictive models accessed via forward rollouts or cost-query interfaces (including simulators and learned world models). SDOB separates mismatch into (1) a state-effect disturbance that changes rollout states and (2) a cost disturbance that changes queried costs at fixed query states, then estimates and compensates both online for improved control performance.

DrivingAI summary
2026-10-06 · arXiv:2610.04911

VideoResearchAgent: Grounded Task Synthesis and Sim-to-Real RL for Open-Web Video Research

Yuhang Zhou, Fei Li, Yuxi Wu et al.

VideoResearchAgent trains agents for open-web video research, where an agent must discover relevant videos, navigate temporal content, and ground answers in visual evidence. The authors add (i) controllable multi-hop video-grounded task synthesis with filtering against text-only solutions, (ii) a field-aligned local video simulator that preserves search/watch interactions and speeds video search, and (iii) Retrieval-Domain-Randomized GRPO to reduce simulator-to-web overfitting. On Video-BrowseComp, Qwen3.5-4B reaches 40.48% OA and uses fewer API tokens.

Video World ModelsCode availableAI summary
2026-10-06 · arXiv:2610.04913

$R^2$-WAM: Repair-and-Reject Post-Training for World Action Models

Ruiyan Xu, Haisheng Su, Sixu Lin et al.

R2-WAM is a two-stage repair-and-reject post-training method for world action models. Predictive repair uses a kinematic alignment score to make imagined futures consistent with the input demonstrated action, then action rejection compares imagined progress of sampled vs demonstrated actions using STEAM to select inferior samples for negative fine-tuning. It improves downstream policy success without extra environment interaction or inference changes.

RoboticsAI summary
2026-10-06 · arXiv:2610.04916

PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation

Jing Xie, Shouwei Ruan, Yubin Wang et al.

PreAct-Nav is a training-free urban navigation harness that equips frozen navigation policies with anticipatory reasoning. It maintains medium-horizon navigation subgoals in Navigation Memory, previews actions via an action-conditioned world model in a Predictive Sandbox, and uses a vision-language model to retain or revise actions before execution, then updates memory from real observations.

DrivingAI summary
2026-10-06 · arXiv:2610.04920

PWM: Personalized World Models with Online Reinforcement Learning

Zhexin Lou, Guancheng Lu, Zeyu Zhang et al.

PWM customizes an action-conditioned interactive world model to a specific scene using a short scene video. It learns a compact LoRA adapter per scene while keeping a frozen Yume-5B backbone, using online reinforcement learning updates from fresh rollouts. PWM is instantiated with GRPO and with DiffusionNFT. A benchmark, PWM-Bench (150 tasks), evaluates on held-out continuations and reports improved customization over native Yume.

Video World ModelsCode availableAI summary
2026-10-06 · arXiv:2610.04940

Software World Models: From Consequence Prediction to Decision Value

Tongli Su, Yuntong Hu, Liang Zhao et al.

The authors introduce the Software World Model (SWM) for coding agents: it predicts a code change’s “blast set” (components the change will break) rather than only the agent’s own observations or static reachability. SWM uses Explore–Learn–Act to collect counterfactual execution outcomes, fine-tune a language-model predictor, and convert predicted blast sets into per-consumer break probabilities for change ranking and proactive migration.

DrivingAI summary
2026-10-06 · arXiv:2610.04988

How Long, Not How Close: A Learned Temporal Metric for Planning in Latent World Models

Lama Moukheiber, Haotian Xue, Yongxin Chen

TEMPO is a temporal-distance planning objective for frozen latent world-model planners (e.g., LeWM, PLDM). Instead of retraining the world model, it learns a small metric on the frozen latent space where distance reflects the number of environment steps between states, and blends this into the planner’s cost. On 5 tasks (1–3 plans from the goal), it raises LeWM on TwoRoom from 36% to 99% for three plans away.

Latent DynamicsAI summary
2026-10-06 · arXiv:2610.05043

CI-JEPA: A Counterfactual Analysis of Latent Representations in Joint-Embedding Predictive Architectures for Self-Supervised Learning

Mintu Dutta, Ritesh Vyas, Mohendra Roy *

CI-JEPA extends I-JEPA with an intervention-aware counterfactual branch that predicts the representation change ΔZ between an original image and a modified one. On Flowers102, using frozen-encoder linear probing, CI-JEPA reports a best validation accuracy of 78.14% vs 77.55% for pretrained ViT-B/16 and I-JEPA. It also reports larger mean L2 representation changes for flower-center occlusion than for background blur/tint.

JEPAAI summary
2026-10-06 · arXiv:2610.05051

LogSig-SSM: Time-Series Modelling with Multi-Scale Log-Signature Compression for State-Space Models

Felix Oury, Nicolas Calvo Peiro, Reiko J. Tanaka

LogSig-SSM compresses long multivariate time series into shorter token sequences by computing multi-scale windowed log-signatures, then models long-range dependencies with a selective SSM backbone (Mamba-style). The authors give a continuous-time (NCDE/NRDE-style) interpretation where selectivity induces input-dependent rescaling of latent dynamics. They report faster training and lower GPU memory than Mamba on the longest sequences across four benchmarks.

Latent DynamicsCode availableAI summary
2026-10-06 · arXiv:2610.05198

Same Predictions, Different Harms: Causal Auditing of Patient World Models

Yicheng Qi, Xiyi Xiong

The authors audit “patient world models” for clinical trial simulation that match transition kernels and arm-specific risks yet may disagree on counterfactual harm (patients harmed by switching treatment). They analyze a two-stage shared-response SCM with categorical intermediate state and common terminal care, deriving sharp harm intervals (exact for up to three intermediate states) and sensitivity bounds under declared stage dependence and response-mismatch budgets. Experiments on public simulators find “honest negative” results for reliability claims.

AI summary
2026-10-06 · arXiv:2610.05240

Pythia: Toward Foundation World Models for Multimodal Time Series

Xilin Dai, Hongzhou Chen, Yifan Hu et al.

Pythia is a context-conditioned time-series foundation world model that learns predictive latent dynamics across datasets using a joint-embedding predictive architecture. It uses a stop-gradient numerical reference to guide contextual corrections in latent space, then freezes the predictive representation and adapts a probabilistic quantile decoder on observed history and predicted world states. On MUSE, Pythia-Tiny reports 0.6879 MASE and 0.4269 WSQL, improving 6.26%/5.00% over the strongest published MUSE baseline.

Latent DynamicsAI summary