Proposes Look-Back, an implicit self-guided mechanism enabling models to spontaneously revisit visual inputs during reasoning, boosting performance by over 7%.

Shuo Yang, Yuwei Niu, Yuyang Liu et al.

HOI-Dyn framework enhances human-object interaction generation using a lightweight Transformer model, reducing object motion inconsistencies.

Lin Wu, Zhixiang Chen, Jianglin Lan

REG method entangles high-level class tokens with low-level latents, boosting diffusion training 63× faster with improved quality.

Ge Wu, Shen Zhang, Ruijing Shi et al.

EvRWKV framework enhances low-light images with continuous interaction, improving PSNR by 1.79 dB and 1.85 dB.

Wenjie Cai, Qingguo Meng, Zhenyu Wang et al.

This study compares MLM and CLM pretraining, finding MLM superior in downstream tasks, but CLM more data-efficient and stable.

Hippolyte Gisserot-Boukhlef, Nicolas Boizard, Manuel Faysse et al.

UniGlyph uses pixel-level text masks and adaptive guidance to improve high-fidelity visual text synthesis in Chinese and English, surpassing prior methods by 15%.

Yuanrui Wang, Cong Han, Yafei Li et al.

This study compares manual, pattern-based, and LLM-generated CQ, revealing LLMs' efficiency but need for refinement.

Reham Alharbi, Valentina Tamma, Terry R. Payne et al.

Proposes a Transformer-based framework for whole-body grasping motion generation, excelling on the GRAB dataset.

Edward Effendy, Kuan-Wei Tseng, Rei Kawakami

World4Drive employs an intention-aware latent world model, reducing L2 error by 18.1%, collision rate by 46.7%, without perception annotations, achieving state-of-the-art results.

Yupeng Zheng, Pengxuan Yang, Zebin Xing et al.

Proposed Inf-Bench framework evaluates spatial deformation reasoning in VLMs, revealing poor 3D task performance.

Jiahuan Zhang, Shunwen Bai, Tianheng Wang et al.

Proposed an efficient depth- and spatially-varying image simulation method, enhancing 12MP real-world deblur performance.

Xinge Yang, Chuong Nguyen, Wenbin Wang et al.

MOVi-MC-AC supplies 5.8M instances across six-camera videos for amodal segmentation, content completion, and cross-view re-identification.

Alexander Moore, Amar Saini, Kylie Cancilla et al.

NavMorph employs a self-evolving world model with continuous latent space and scene memory, boosting VLN-CE performance and online adaptation.

Xuan Yao, Junyu Gao, Changsheng Xu

OmniVCus employs a diffusion Transformer with multimodal control, enabling multi-subject video customization without labels, using LE and TAE mechanisms.

Yuanhao Cai, He Zhang, Xi Chen et al.

RoboTwin dual-arm challenge, using SEM and AnchorDP3, achieved over 97% success in simulation.

Tianxing Chen, Kaixuan Wang, Zhaohui Yang et al.

VolumetricSMPL achieves 10x faster inference using Neural Blend Weights.

Marko Mihajlovic, Siwei Zhang, Gen Li et al.

AlignEvoSkill enhances skill evolution via knowledge tags and task alignment, achieving a 34.7% improvement.

Dingzirui Wang, Xuanliang Zhang, Keyan Xu et al.

HERB benchmark reveals retrieval bottlenecks in RAG systems over heterogeneous enterprise data, scoring 32.96 on average.

Prafulla Kumar Choubey, Xiangyu Peng, Shilpa Bhagavath et al.

Proposes multimodal perception and world modeling framework for autonomous embodied AI, achieving 85% action prediction accuracy.

Pascale Fung, Yoram Bachrach, Asli Celikyilmaz et al.

Defines scoring bias, identifies three types—Rubric order, Score ID, reference answer bias—and proposes a quantification framework.

Qingquan Li, Shaoyu Dou, Kailai Shao et al.