arXiv Machine Learning

Contrastive World Models

Contrastive World Models propose a new method for learning latent dynamics without pixel reconstruction. By replacing observation reconstruction with a Deep InfoMax-like objective that maximizes mutual information between state-action sequences and local patch features of future observations, the approach encourages state representations to retain predictive information while ignoring visually irrelevant details. Experiments show that this method matches existing baselines in simple settings and significantly outperforms them when distractors or natural video backgrounds are present, while also training more efficiently by eliminating the pixel decoder.

arXiv Computer Vision
Sep 28

StarWM: Self-Supervised Trained Attention Routing for Robust World Models

StarWM introduces a self‑supervised attention routing mechanism that selectively applies reconstruction only to dynamically relevant regions of visual input. By combining a cross‑attention module with a dual‑stream decoder and stop‑gradient barriers, it balances faithful environmental dynamics capture with abstraction of irrelevant content. Experiments on DeepMind Control show that StarWM outperforms both reconstruction‑based and reconstruction‑free baselines, especially under distractor conditions, and preserves state attributes over long‑horizon imagination.

By Zeqiang Zhang, Fabian Wurzberger, Maximilian Otte, Daniel Schmid, Sebastian Gottwald, Arne Peter Raulf, Daniel Alexander Braun
arXiv AI
Nov 6

MAC-v2: Long-horizon model-based value expansion from pixels

MAC‑v2 is a new framework for long‑horizon model‑based value expansion from pixel observations in offline reinforcement learning. It trains a reconstruction‑free latent world model that learns representations through prediction and control objectives, avoiding the need to encode control‑irrelevant visual details. By using action chunking, MAC‑v2 reduces recursive model predictions, mitigating error accumulation and achieving strong performance on visual benchmarks such as OGBench, MMBench2, Robomimic, and LIBERO‑10.

By Kwanyoung Park, Sergey Levine
arXiv Computer Vision
Nov 5

Motion-guided Contrastive Regularization For Latent World Models

The paper introduces motion-guided contrastive regularization to improve JEPA-style world models by preventing static collapse, where encoders overemphasize episode-specific static content. By comparing predicted latent transitions with transitions to candidate observations within the same episode and using optical-flow similarities as soft contrastive targets, the method encourages representations that preserve manipulation-relevant state changes. Experiments on OGBench-Cube and LIBERO-Object show significant gains in planning success while maintaining performance on other visual-control benchmarks.

By Yiwen Chen, Qiqi Hou, Jisoo Jeong, Yinhao Zhu, Huaizu Jiang, Amir Said
arXiv AI
Aug 10

TaskSense: Focusing on What Matters in World Models

arXiv:2608. 06544v1 Announce Type: new Abstract: World models for visual control typically learn compact latent states by reconstructing observations, implicitly encouraging representations to preserve information across the entire visual input.

By SM Mazharul Islam, Manfred Huber
arXiv AI
Oct 28

Vision-Language Agents Can Learn in Their Own Latent Worlds

The paper introduces Dwell, a model‑based reinforcement learning framework that treats a pretrained vision‑language model (VLM) as a recurrent latent transition operator. Dwell enables the VLM to predict, adjust, and act within its own latent space, allowing the agent to learn from real interactions and latent imagination without reconstructing full futures. Experiments across five interactive visual environments show that Dwell outperforms other trainable methods while significantly reducing computational demands for imagination.

By Haoxi Li, Sikai Bai, Jianfei Ma, Qinglin Hou, Xiaocheng Lu, Tao Han, Jie Zhang, Song Guo
arXiv AI
Jun 16

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

arXiv:2606. 15768v1 Announce Type: cross Abstract: Vision-Language-Action models (VLAs) leverage large-scale vision-language pretraining for semantic robot control, but often lack explicit foresight into how robot actions change the scene.

By Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, Yuanbo Xu, Chao Yu
arXiv AI
Jul 1

Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding

arXiv:2606. 31232v1 Announce Type: new Abstract: Learning visual world models for planning requires compact latent dynamics that remain sensitive to actions, yet reconstruction-free joint-embedding objectives can collapse to action-insensitive representations.

By Zhenghao Zhang, Yuanxiang Wang, Zhenyu Guan, Yujia Yang, Bingkang Shi, Tianyu Zong, Hongzhu Yi, Guoqing Chao, Xingchen Chen, Tiankun Yang, Chenxi Bao, Tao Yu, Jingjing Zhou, Jungang Xu