Digest

Reset

1167 papers

2026-10-09 · arXiv:2610.11175

Higher-Order Action Supervision Makes A Strong Policy Class

Peng Cheng, Yunxian Hou, Zhi Zhou et al.

The authors argue control instability and poor robustness in offline learning come from supervising only 0th-order actions (action labels) and ignoring higher-order action dynamics/temporal consistency. They propose a plug-and-play loss that adds first-order (time-derivative) action supervision to off-the-shelf policy models (deterministic, stochastic, or flow) without structural changes, with theoretical guarantees (IL) and improved performance/robustness on OGBench and D4RL.

DrivingAI summary
2026-10-09 · arXiv:2610.11248

SimVLA: Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation

Kyoungin Baik, Youngwoon Lee

SimVLA is an end-to-end, zero-shot sim-to-real VLA framework for mobile manipulation trained entirely on synthetic simulation data without teleoperation. It pre-trains on SimAction (35 tasks via skill composition) and SimVQA (privileged simulator-derived VQA supervision), then post-trains on a mixture of SimAction and SimDeploy (policy rollouts). The authors report zero-shot transfer to real environments and better performance than a π0.5 fine-tuned with 50 in-domain real demonstrations.

DrivingAI summary
2026-10-09 · arXiv:2610.11253

LLM-IDEA: Identifiability-Driven Experimental Agent for Autonomous Discovery of Mechanistic World Models

Surya Shetty, Ulisses Braga-Neto

LLM-IDEA is a closed-loop autonomous discovery agent for mechanistic world models that computes identifiability from the agent’s own proposed model and data. It outputs a three-way plateau verdict—capability limit, resolvable within the design class, or certified exhausted—and uses this to decide whether to search more, run better same-kind experiments, or change the experiment kind. Benchmarks include ODEBench, DiscoverPhysics, and an Alien Universe testbed.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.11283

Being-M0.7: A Latent World-Action Model for Humanoid Robots

Junpeng Yue, Boyuan Li, Yuxuan Wang et al.

Being-M0.7 is a latent world-action model for humanoid loco-manipulation that transfers visual–motion priors from mixed-modality human video and motion data to robot control. It uses a Mixture of Transformers prior pretrained on latent visual-state and motion prediction, adapted by robot mid-training, then grounded into executable actions via a gated action expert post-training. The authors report the highest aggregate simulation success on SIMPLE and matching the strongest baseline on real-world Unitree G1 tasks.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.11376

CRISP: Fixing Flying Pixels in Latent LiDAR Generation via Diffusion Decoding

Andrea Ceron, Michael Schmidt, Alvaro Marcos-Ramiro et al.

CRISP replaces the decoder in latent LiDAR pipelines with a pixel-space diffusion decoder to fix “flying pixels” from convolutional VAE decoders blurring depth discontinuities. It uses a backbone-agnostic latent adapter, a DiT denoiser, and a support mask predictor, while keeping the encoder/latent generator fixed. The authors report large reductions in FSVD/FPVD/FRID and improvements in edge geometry, including zero-shot decoder swaps inside a pretrained LiDM world model.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.11382

PlanWAM: Planning-Shaped Future Representations for End-to-End Autonomous Driving

Jinchang Xu, Hongda Yu, Fengwei Dong et al.

PlanWAM is a planning-shaped world action model for end-to-end autonomous driving. It learns a compact recency-aware history representation, then uses a training-only privileged posterior (ground-truth future frames) to shape a future latent via trajectory generation/selection objectives. A prior branch distilled from the posterior predicts this planning-shaped future latent from history at inference, providing foresight for trajectory planning.

DrivingAI summary
2026-10-09 · arXiv:2610.11401

WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models

Kai Ding, Yang He, Ruijie Quan et al.

WAM-Cache is a training-free KV-reuse method for World Action Models that keeps video DiT layerwise key–value pairs across inference chunks and recomputes only a sparse refresh set selected by (i) latent surprise, (ii) action-conditioned attention, and (iii) an age bound. On Fast-WAM, it cuts video DiT prefill FLOPs by 32–42% while staying close to dense policy success.

DrivingAI summary
2026-10-09 · arXiv:2610.11416

Rewiring Semantics, Dynamics, and Control: A Simple yet Effective Action-Centric Tri-Stream Transformer

Shuang Luo, Yilun Kong, Yunpeng Qing et al.

ACT^3 is an action-centric tri-stream transformer for vision-language-action control that fuses semantic understanding (from a VLM backbone) and predictive dynamics (from a video-generation WM) through an action expert. Semantic and dynamics context streams run independently; only the action queries attend to both via layerwise cross-stream attention, and both backbones are updated via action supervision.

RoboticsAI summary
2026-10-09 · arXiv:2610.11444

Learning to Retrieve: Internalizing Memory Retrieval for Video World Models

JiaKui Hu, Tailai Chen, Yuqi Pan et al.

The authors propose Learning-to-Retrieve (L2R), which internalizes memory retrieval inside a video diffusion transformer for camera-conditioned video world models. L2R repurposes a persistent internal state as memory, using a camera-conditioned retrieval gate (what to retrieve) and a retrieval trigger supervised by a 3D re-visibility signal (when to retrieve). This aims to improve long-term scene consistency without an external memory bank or 3D conditions.

2026-10-09 · arXiv:2610.11483

BridgeGuard: Explicit Safety Drift for Diffusion-based Autonomous Driving

Zhenjun Qiu, Jianing Huang, Dongang Liu et al.

BridgeGuard is a safety-constrained diffusion planning method for autonomous driving. It progressively strengthens a constraint during denoising to move intermediate trajectories toward a scene-dependent safety domain. It learns a time-dependent obstacle distance field (DistanceFieldNet) from BEV features using value and spatial-gradient supervision beyond expert trajectories, and injects the learned safety correction via low-dimensional curve updates while freezing the pretrained perception backbone and planner.

DrivingAI summary
2026-10-09 · arXiv:2610.11527

AtomWorld-Mirror: Macro-Step World Modeling of Critical Evolution Backbones for Materials Dynamics

Ziming Pan, Ruge Zhang, Haozhi Han et al.

AtomWorld-Mirror is a time-aware macro-step world model for long-horizon atomistic materials evolution. It distills short micro-event segments into physically reachable transitions between key “critical evolution backbone” states, jointly predicting sparse structural edits and accumulated continuous-time. The model enforces local reachability, inventory conservation, and continuous-time consistency, and replaces explicit micro-event replay with macro-step inference for large speedups.

Latent DynamicsAI summary
2026-10-09 · arXiv:2610.11552

Safe, Persistent, and Evolving Agent Harness for Understanding Partially Observable Worlds

Yisen Gao, Yue Guo, Qing Zong et al.

The authors introduce E-LEDGER, a multi-agent harness for safe, persistent execution in partially observable enterprise workflows. It compiles policies into executable code approvals, and maintains a world ledger with verified hidden rules plus evidence-backed dynamic state. They further propose WorldAbduct, which uses four diagnostic views and targeted abduction interactions to recover hidden rules for the ledger. On WOW, E-LEDGER+WorldAbduct improves STCR by outperforming the strongest evolution baseline by 5–15 percentage points.

Code availableAI summary
2026-10-09 · arXiv:2610.11591

Acting from Belief, Looking When Needed: A Bayesian Spatial World Model for Navigation under Intermittent Perception

Feihong Yang, Xiang Long, Jincheng Yu et al.

ALONE is a Bayesian spatial world model for navigation under intermittent perception. It propagates a structured spatial belief between depth observations and requests a new observation only when a reliability map indicates insufficient support for action selection and the relevant region is expected to be visible. On drone navigation with intermittent single-camera depth, it reports 98%/97% closed-loop success at 10 Hz, with median depth requests at 0.9% and 1.3% of decision steps (successful trials), and all 10 real-world trials successful.

DrivingAI summary
2026-10-09 · arXiv:2610.11617

TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction

Yuqi Li, Xiaoqin Feng, Fan Xu et al.

TAM (Task-Aware Memory Distillation) organizes a frozen teacher’s knowledge into a bounded, retrievable history to provide cross-sample supervision for spatiotemporal prediction. Task-specific rules choose what representations to store (latent features, forecast changes, or flow residuals) and which historical entries to retrieve. The student learns either relational matching over shared references or residual-prototype regression; teacher/memory/adapters are training-only, so inference cost is unchanged.

2026-10-09 · arXiv:2610.11692

Can Jev be Your Q or Policy in Reinforcement Learning?

Yi Ma, Tianpei Yang, Yaodong Yang et al.

The authors study whether Jev, a frozen decision model that returns calibrated typed answers in a single forward pass, can serve as Q-like or policy-like components inside RL training. They formalize Jev’s usable roles (policy, score rater, meta regulator, transition; excluding cardinal value arithmetic) and propose algorithms using Jev as reference policy, exploration judge, and replay rater. Across nine MiniGrid tasks and three Atari games, Jev training outperforms a standard RL learner.

DrivingAI summary
2026-10-09 · arXiv:2610.11704

Intervention anchors and scientific verification in synthetic vascular predictive representations

Lingsen You, Yujun Guo, Xinyu Zhong et al.

The authors present a synthetic, mathematical audit of when orthogonal predictive coordinates can be interpreted as named intervention edits. In a controlled linear setting, prediction is preserved under coordinate transforms but named edits require anchors and correct coordinate transport. They report near-machine-precision prediction equivalence yet nontrivial median edit errors (median 1.513) without calibration; coordinate transport removes that error. They also run a fault-bank with executable verification contracts for acceptance/rejection of supported edits.

DrivingAI summary
2026-10-09 · arXiv:2610.11723

MultiWorldBench: Do Independently Controlled Views Describe One Shared World?

Zhangbo Xu, Ruoxi Zhang, Rui Hu et al.

The authors introduce MultiWorldBench, a Minecraft diagnostic benchmark (495 cases) for multiplayer world models: two independently controlled views should reflect one shared, persistent world. They evaluate Solaris, Gamma-World, and MineWorld vs Engine GT. Gamma-World ranks highest among generated systems (21.39 ten-capability average), but all generated systems score very low on state persistence and structural consistency, and none succeeds on spatial reasoning or building-identity preservation.

DrivingAI summary
2026-10-09 · arXiv:2610.11794

Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks

Haoyu Zhao, Zhengxu Yu, Zhiyuan He et al.

Memento 3 (“Code as Model”) uses a frozen LLM with external memory to learn explicit world models for planning in unfamiliar environments. The agent keeps a natural-language rulebook (semantic memory), compiles it into executable prediction code, and iteratively updates both via observation–reflection–rule revision–compilation–verification. Verified updates guide further interaction; the LLM parameters remain fixed.

DrivingAI summary
2026-10-09 · arXiv:2610.11942

Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents

Jiaming Zhang, Xuan Wang, Fuyao Zhang et al.

For GUI agents, safety depends on whether the outcome matches the expected transition, not on whether the resulting screen looks legitimate. The authors argue that verification should predict the next observation representation directly (decoder-free) and compare vectors. They present LGWM, trained on 1.85M real GUI transitions, and evaluate on RSWT-Bench where outcome-only detectors are at chance.

DrivingAI summary
2026-10-09 · arXiv:2610.11975

Predicting Cable Dynamics with Physical Attention Bias

Avihai Giuili, Rotem Atari, Avishai Sintov et al.

The authors study learned simulators for deformable linear objects (cables) and ask what geometry should bias global attention between distant cable segments. They add a physical attention bias to attention logits using either arc-length (intrinsic) distances, Euclidean (extrinsic) distances, or both on disjoint attention heads. Physical bias improves prediction on unseen cables; arc-length yields the largest gain when attention is the only long-range coupling. Code: https://github.com/avihaig/dlogps.

DrivingCode availableAI summary