SmolWAM is a compact latent world‑action model designed for robotic manipulation on edge devices. It is derived from the 2B‑parameter Cosmos Policy by reusing its pretrained trunk and selecting 12 of 28 transformer blocks, while employing single‑frame visual latents and cached text features to produce actions and future visual latents in a single rectified‑flow step. Trained in only 3.6 GPU‑hours, SmolWAM achieves 95.5% success on 400 LIBERO episodes at 57 ms per step, using 2.3 TFLOP per 16‑action chunk and 2.4 GiB GPU memory—an 11× speed‑up, 18.5× fewer FLOPs, and 3.2× less memory than its teacher model, while still forecasting scene motion.
By Timi Owolabi, Hikmah Olawore, Frances Adelakun, Ayotomiwa Oyewumi
The paper introduces Semantic Head Specialization (SHS), a phenomenon where Vision Transformer (ViT) attention heads specialize as either object- or background-focused, most evident under full attention. It proposes the SHS-Index to quantify this specialization, demonstrating its ability to distinguish full-attention from chunk-window ViTs and its strong correlation with downstream benchmark performance. Leveraging insights into window interaction, token serialization, and local softmax allocation, the authors design Ariadne Attention, a hybrid attention mechanism that matches full-attention performance on 22 image and video tasks while reducing attention compute by 6.5×.
By Chenhong He, Lei Li, Shicheng Li, Hanglong Lv, Lingpeng Kong, Qi Liu, Tong Yang, Shuhuai Ren
arXiv:2610.31762v1 Announce Type: cross
Abstract: What visual representation should a navigation policy act on? Navigation policies built on vision foundation models largely leave representation desi...
By Jiawei Chen, Yinxi Wang, Zihan Yang, Meijun Gao, Yaowei Liu, Shuming Zhang, Zecheng Li, Xiangju Mi, Yuxuan Liu, Runzhong Huang, Tianlin Li, Xun Zhang, Zhengyang Wang, Xu Miao, Tong Yang
PhysBrain 1.5 is a unified vision‑language model that learns to understand physical environments, generate actions, and predict future states by encoding language, end‑effector motion, and dense visual targets as discrete sequences and training them with autoregressive next‑token prediction. The model is pre‑trained on human interaction videos and fine‑tuned on human demonstrations, robot trajectories, and simulated experience, achieving an average score of 72.5 across 28 embodied understanding benchmarks and outperforming other open‑source models on 14 of them. It also demonstrates the ability to produce end‑effector trajectories and predict future scenes with spatially aligned RGB, depth, and robot‑mask outputs.
By DeepCybo Team, Yu Bin, Haipeng Cao, Zheng Chang, Kai Chen, Youning Chen, Kailin Deng, Yichao Du, Xiaotong Fu, Haoyang Ge, Yunlong Guo, Chenliu Hao, Jiyan He, Xuguo He, Yakun Hou, Kai Hu, Cong Huang, Tuopusen Huang, Yu Huang, Hong Li, Peize Li, Shijie Lian, Xiaopeng Lin, Yun Lin, Haibao Liu, Haochen Liu, Qiuzhi Liu, Shengcai Liu, Zhiqiang Liu, Tao Luo, Peng Ren, Shuo Ren, Chaoyi Ruan, Zhaolong Shen, Yukun Shi, Qiyuan Su, Yuxuan Tian, Yining Wang, Changti Wu, Hao Wu, Xueyin Xu, Ruoqi Yang, Zhaoyang Yang, Hang Yuan, Zhaoyang Zeng, Hanwen Zhang, Ruimeng Zhang, Yao Zhang, Yibo Zhang, Yuxiang Zhang, Zhirui Zhang, Ziyi Zhang, Zubin Zheng, Zishen Zhuang
We introduce Predictive State Retrieval (PSR), a task in which a model observes a short video prefix and a temporal question about an object's future state, then retrieves instances from other videos or images that depict that state. Unlike action anticipation, which predicts a label, moment retrieval, which localizes an observed event within a video, or video generation, which synthesizes pixels, PSR combines anticipation with cross-instance retrieval across multiple temporal horizons.
The paper introduces 4DMemory, a benchmark for long‑horizon spatiotemporal memory that includes the LongVideo dataset—about 900 hours of video with over 5,000 manually reviewed questions—and the 4DMemoryBench. It also presents the 4DMemory Harness, a system that combines four complementary skills, more than 22 specialized tool models, and a retrieval‑augmented generation component. Evaluation of 13 leading models shows that standalone foundation models perform poorly on long‑horizon tasks, while the 4DMemory Harness boosts performance by over 50% and reveals a performance drop of more than 10% compared to previous benchmarks.
By Yanhao Jia, Yi Zhu, Yiran Zhong, Weigao Sun, Feida Zhu, Pengkun Jiao, Sihan Yang, Ming Ma, Steven Hoi, Erik Cambria
MT‑WAM enhances the Fast‑WAM framework by adding complementary supervision for future 2‑D point trajectories and visual features while keeping the original training objectives. A lightweight dual‑stream branch and structured attention mask isolate motion‑specific processing, and motion‑stream tokens provide additional dynamics cues to the action expert. During inference, MT‑WAM skips future‑video prediction, using cached video and motion information to achieve higher success rates on LIBERO, LIBERO‑Plus, RoboTwin 2.0 Clean2Rand, and several real‑world tasks.
By Yiguang Yang, Jiankun Peng, Xiaoming Wang, Yiran Zhang, Zhibo Fang
arXiv:2609.32193v2 Announce Type: replace
Abstract: Vision Language Action (VLA) models condition actions directly on current visual and language context, without an explicit account of how the scene...
By Hongyi Cai, Yi Herng Ong, Tingshiuan C. Wu, Chiew Hui Lim, Hanxia Li, Kehong Guo, Sze Yuan Cheong
The paper introduces Gated Token Recurrence (GTR), a softmax‑free recurrent vision backbone that replaces global softmax attention with gated linear attention, alternating scan directions, and enhanced SwiGLU blocks. GTR is distilled from a DINOv3 teacher using only final‑layer patch‑token alignment, and achieves strong performance on COCO object detection (58.9 box AP) with very low latency (1.908 ms on an RTX 4090). The backbone also transfers to multiple dense prediction tasks and runs efficiently on edge hardware via a specialized CUDA operator and TensorRT deployment.
By Zhe Feng, Longfei Liu, Wei Liu, Kai Chen, Jiangjiang Kong, Wei Zhou, Yifeng Qian, Dexiong Chen, Xuanlong Yu, Xi Shen
The paper introduces the Just Action Model (JAM), a framework that decouples backbone initialization, future observation objectives, and inference machinery to unify vision-language-action (VLA) and world-action model (WAM) approaches. JAM employs a pretrained visual backbone, an action expert, an action loss, and a foresight loss that predicts future encodings without visual generation, discarding the prediction head after training. Experiments show JAM outperforms FastWAM and ImageWAM on LIBERO, LIBERO-Plus, and RoboTwin, while reducing online training FLOPs by 73%.
By Yonghao Yu, Lang Huang, Runyi Li, Toshihiko Yamasaki
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:2610.27851v1 Announce Type: cross
Abstract: This paper shows that JEPA-style latent world models (LWMs) can learn across diverse domains and data sources by adopting the language-conditioned in...
By Junha Song, Geonmo Gu, Jeongeun Park, Heejeong Nam, Lucas Maes, Byeongho Heo, Dongyoon Han, Sangdoo Yun