Digest

Reset

1167 papers

2026-10-14 · arXiv:2610.18868

Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards

Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh et al.

Tev proposes Tev-Gym, a long-horizon agent training dataset with verifiable, dense, progressive rewards built from a latent-world-first synthesis approach (latent specification → progressive rendering with exact verification). It trains models with progress-potential reward shaping (PPRS), yielding Tev-27B that matches GPT-6 Astra and Claude Opus 5 at Pass@5 on ALE-CM, and generalizes to Terminal-Bench (TB) subsets.

Latent DynamicsAI summary
2026-10-14 · arXiv:2610.18873

LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts

Wenxuan Guo, Yingping Liang, Zhaoguo Wang et al.

LoopWAM is a closed-loop world-action model that learns directly from mixed-quality robot deployment rollouts (successes, suboptimal behavior, and failures). It uses source-dependent noise sampling, quality steering, and quality-weighted action supervision while training video and action tokens together through parallel denoising. On RoboChallenge Table30-V2 it improves success from 23.67% to 43.33% across three deployment-training rounds.

DrivingAI summary
2026-10-12 · arXiv:2610.12567

CASTER: Formalizing Zero-Shot Manipulation from Synthetic Videos as Constrained Optimization

Yinlong Dai, Homanga Bharadhwaj, Dylan P. Losey

CASTER formalizes zero-shot manipulation from synthetic videos as constrained optimization: synthetic videos seed an initial guess of object motion, and a robot refines trajectories in a digital twin to satisfy semantic consistency and physical feasibility. The authors analyze requirements on video generation via “semantic basins of attraction” and report improved average zero-shot task success (76.6%) over the best-performing baseline by 34.1%.

RoboticsAI summary
2026-10-12 · arXiv:2610.12572

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang et al.

DiffusionART fine-tunes diffusion models with black-box rewards in an efficient and stable way. The authors diagnose instability in NFT-style forward-process RL as a hidden forward–reverse consistency force and argue EMA stabilizes via restoration but needs per-task tuning. DiffusionART aggregates each prompt’s rewards into a shared reward-improving direction for reverse states and uses a trust-region-masked two-update restoration, improving stability without EMA tuning.

Video World ModelsCode availableAI summary
2026-10-12 · arXiv:2610.12575

Mind the Temporal Gap: Evaluating Video Models for Temporal Understanding in Driving

Shiva Sreeram, Erfan Aasi, Tsun-Hsuan Wang et al.

The authors evaluate whether video-language models reliably infer ego-motion from driving videos. They find standard, visual, few-shot, and chain-of-thought prompting are unreliable and highly model/task dependent, with persistent class biases. In contrast, lightweight LoRA finetuning and simple probes on intermediate visual features recover strong motion-discriminative performance, suggesting motion cues exist internally but are not reliably used by the video-to-language pathway.

DrivingAI summary
2026-10-12 · arXiv:2610.12696

Spa-Bench: A Real-Robot Benchmark for Spatial Grounding in Vision-Language-Action Models

Justin Tien-Smith, Matthew Pidden, Rolandos Alexandros Potamias

Spa-Bench is a real-robot benchmark for spatial grounding in vision-language-action models. It evaluates six tabletop tasks (object state, relative size, referential description, relational placement, ordinal position, counting) under in-distribution vs out-of-distribution spatial configurations using a withheld-cell split and four diagnostics. On 5 evaluated policies, success drops on OOD spatial configurations, indicating weak robust spatial grounding.

DrivingAI summary
2026-10-12 · arXiv:2610.12738

KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting

Jonas Petersen, Pratik Jawahar, Riccardo Maggioni et al.

KITE (Knowledge-Injected Temporal Encoder) injects engineering “semantic priors” into a self-supervised horizon-conditioned event forecaster using two switchable mechanisms: a covariate route and an attention-bias route (each identity at its null). The authors report that informative priors help in short contexts and out-of-domain transfer by providing information the observed window lacks, while redundant priors and the graph bias show no measurable gain.

Latent DynamicsAI summary
2026-10-12 · arXiv:2610.12789

EvoBelief: Falsifiable World Models for Self-Improving LLM Agents

Janet Wang, Jiang Guo, Yi Xiang et al.

EvoBelief is a self-evolving world model for frozen-LLM agents that represents environment knowledge as falsifiable hypotheses. During deployment, mismatches or decision-critical unknowns trigger hypothesis formation; later observations or controlled experiments validate them before they guide predictions/plans. Counterexamples narrow or retire rules, and validated rules are reused across tasks.

DrivingAI summary
2026-10-12 · arXiv:2610.12818

LeWAM: End-to-End World and Action Modeling with JEPAs

Minghao Fu, Tavis Siebert, Eryk Halicki et al.

LeWAM is an end-to-end joint-embedding predictive architecture for control that learns (i) next latent-state prediction and (ii) action flow matching with an isotropic Gaussian latent constraint (SIGReg). A shared Mixture-of-Transformers predicts latent dynamics and learns a stochastic action policy, enabling direct policy execution and faster visual planning via gradient-based refinement through learned latent dynamics.

Latent DynamicsAI summary
2026-10-12 · arXiv:2610.12847

When Generative Models Act as Simulators

Sevan Brodjian, Michael Hobley, Pietro Perona

The authors test when a generative model can act like a simulator via “editability”: after learning from observations only, can they map a simulator state edit into the model’s latent state so the next-step output matches the simulator’s edited rendering? They compare probe-based representation steering vs direct inverse mapping in Othello and Rayworld, finding environment-dependent success: steering degrades under rule simplification in Othello, while probes fail in Rayworld’s continuous setting yet inverse maps can still edit successfully.

Latent DynamicsAI summary
2026-10-12 · arXiv:2610.12881

Mind the Spatial Gap: A Benchmark for Geospatial Reasoning and Constrained Spatial Editing

Zheng Ren, Yanning Wang, Ziheng Huang et al.

GeoSpatialBench evaluates spatial relation understanding and constrained spatial editing on real urban scenes. Using 28 formal relations across GeoJSON and rendered map images (and Together), the authors find a “spatial gap”: GeoJSON generally outperforms map images, Together does not consistently improve, and editing—especially for ordering—often fails due to constraint violations.

DrivingAI summary
2026-10-12 · arXiv:2610.12886

LeapBot-JAC: Joint Action Chains for World Action Models

Chenhao Zhu, Baibing Liu, Pei Liu et al.

LeapBot-JAC addresses action-chunking policies that predict future controls but execute only a prefix before replanning. The authors propose LeapBot-JAC, which reorganizes a fixed action-token budget into a sparse reference chain and a dense execution chain within one shared action expert. Both chains are supervised by demonstrated controls and jointly denoised with aligned physical-time positions; directed attention lets execution tokens read reference tokens under the host’s visibility constraints.

DrivingCode availableAI summary
2026-10-12 · arXiv:2610.12910

Sampling-based Kinodynamic Planning with World Models

Yitian Gao, Hanlan Yang, Zachary Kingston

The authors propose kinodynamic sampling-based planning with a frozen visual world model (DINO-WM) used as a latent-space propagator. They replace myopic CEM/MPPI with latent tree search (nearest-neighbor-guided RRT or projection-based KPIECE) plus short CEM “polishing,” using GPU-native batched expansion. They evaluate on four simulated domains and a real block-pushing task.

DrivingAI summary
2026-10-12 · arXiv:2610.12936

Evo-World: A Lightweight Vision-Language-Action Model with Future Motion Prediction via Optical-Flow Supervision

Zile Zhong, Yuxin Du, Tao Lin et al.

Evo-World is a lightweight vision-language-action model that augments policy learning with future motion prediction via optical-flow supervision. It trains a motion representation module to predict task-relevant motion in latent space, using optical-flow renderings only during training, and preserves these predictive representations during policy learning via SIGReg and detached gradient paths. It reports strong simulation success and better real-world manipulation performance on an AgileX dual-arm robot while using 0.87B parameters.

DrivingCode availableAI summary
2026-10-12 · arXiv:2610.13020

Trajectory-Aware Policy Optimization: Closing the State Distribution Gap in Long-Horizon LLM RL

Zhicheng Cai, Hanlin Wu, Weinan Dai et al.

The authors identify a “state distribution gap” in long-horizon autoregressive LLM RL: token-level policy mismatch compounds into trajectory-induced state distribution shift, causing training instability and eventual collapse. They propose Trajectory-Aware Policy Optimization (TAPO), which derives a state-distribution correction from the off-policy policy gradient, uses geometric-mean trajectory correction for variance reduction, and applies length-aware dynamic clipping. Experiments on eight benchmarks show improved performance and stability over existing LLM RL algorithms.

DrivingAI summary
2026-10-12 · arXiv:2610.13021

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

Jie Huang, Yanan Chen, Ruixun Liu et al.

EmbodiedCLBench is introduced as “the first benchmark in embodied simulation” to evaluate continual learning for self-evolving agents. It uses compositional generalization: agents first learn basic embodied tasks, then are tested on unseen advanced tasks requiring recombination of learned skills. Experiments across multiple harnesses/models show advanced tasks are hard in zero-shot, but basic-task experience yields consistent improvement that stays stable with complexity yet does not scale with more learning samples.

DrivingAI summary
2026-10-12 · arXiv:2610.13024

What Visual Representations Do World Action Models Need?

Kai Liu, Jiangran Lyu, Tingyu Deng et al.

World Action Models (WAMs) use VAE representations learned for visual reconstruction rather than manipulation. The authors run controlled comparisons of visual representations, five feature-injection mechanisms, and feature combinations, and instantiate the best “Representation Recipe” as R2-WAM. They report robustness gains from general-purpose representations, highest success with channel-wise joint prediction, and limited additional gains from composing more sources beyond the first improvement.

Latent DynamicsAI summary
2026-10-12 · arXiv:2610.13117

Forward Dynamics Models

Zipei Ma, Junzhe Jiang, Jiahui Zhang et al.

Forward Dynamics Models (FDM) couple action learning with future world-state prediction using an action-to-future paradigm: future prediction reads action features, while action prediction cannot access future information. End-to-end training uses ground-truth future states to supervise the policy through action features, so prediction errors cannot propagate into actions. At inference, FDM skips future generation, and is instantiated with both VLM and VGM backbones.

DrivingAI summary
2026-10-12 · arXiv:2610.13164

Grounded-Dreamer: Robot Learning from World Model Synthetic Data with Physical Grounding

Feng Jiang, Shu Chen, Yuanfei Wang et al.

Grounded-Dreamer synthesizes physically grounded robot manipulation data using a video world model plus simulator physics. Starting from a small set of real images of task initial states, it reconstructs task-relevant objects and camera geometry, collects grounded trajectories in simulation, post-trains a video world model and an inverse dynamics model, generates manipulation videos from real images, and filters rollouts via closed-loop simulation replay (keeps only task-successful). Policies are trained only on this grounded synthetic data.

2026-10-12 · arXiv:2610.13172

RoleWorld: Knowledge Transfer in World Modeling with Functional Role Abstractions

Boyuan Xiao, Pu Wang, Ji Feng et al.

RoleWorld is a symbolic world modeling approach for knowledge transfer. A teacher LLM extracts reusable role prototypes and executable transition rules, plus a role-guided skill from source trajectories. In a target world without the teacher, the actor keeps role prototypes/rules and the skill fixed, refines object–role assignments and environment-specific parameter estimates online, and uses active verification when uncertainty matters. Experiments on MARS report robust generalization across environment variants with altered object behavior.

Latent DynamicsAI summary