Depth Anything V2 leverages synthetic images, larger teacher models, and large-scale pseudo-labels to achieve finer, more robust monocular depth estimation, over 10x faster than prior models.

Lihe Yang, Bingyi Kang, Zilong Huang et al.

Proposes Sketchpad, enabling multimodal models to draw visual sketches, boosting math and visual reasoning by over 12%.

Yushi Hu, Weijia Shi, Xingyu Fu et al.

Yo’LLaVA employs learnable latent tokens for few-shot personalization, significantly enhancing personalized vision-language understanding.

Thao Nguyen, Haotian Liu, Yuheng Li et al.

OmniTokenizer is a unified image-video tokenizer using decoupled spatial-temporal transformer architecture, achieving SOTA reconstruction with 1.11 FID on ImageNet.

Junke Wang, Yi Jiang, Zehuan Yuan et al.

LVNet introduces a hierarchical keyframe selector (HKS) for efficient long-video QA, achieving 61.1% accuracy with only 12 frames, without training.

Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya et al.

Proposes dynamic oracle-guided autoregressive training, improving NER and summarization, with exact and approximate algorithms for specific metrics.

Jianing Yang, Harshine Visvanathan, Yilin Wang et al.

MLKV introduces cross-layer KV sharing, reducing memory cache by 6× with minimal performance loss in Transformer decoding.

Zayd Muhammad Kawakibi Zuhri, Muhammad Farid Adilazuarda, Ayu Purwarianti et al.

Proposes Softmax-DPO (S-DPO) to enhance recommendation by leveraging multi-negative samples and softmax ranking, significantly outperforming baselines.

Yuxin Chen, Junfei Tan, An Zhang et al.

Proposes AdaSlot, a dynamic slot attention framework that adjusts object slot numbers based on scene complexity, improving unsupervised object discovery.

Ke Fan, Zechen Bai, Tianjun Xiao et al.

Proposes a human-centric digital system engineering framework integrating proactive, reactive, and passive roles, emphasizing trust and regulatory compliance.

Martina De Sanctis, Paola Inverardi, Patrizio Pelliccione

EgoExo-Fitness dataset combines synchronized egocentric and exocentric videos, enabling multi-task full-body action understanding with rich annotations.

Yuan-Ming Li, Wei-Jin Huang, An-Lan Wang et al.

OmniH2O employs goal-conditioned imitation learning and multimodal interfaces to enable dexterous whole-body humanoid teleoperation and autonomous control.

Tairan He, Zhengyi Luo, Xialin He et al.

Proposes Center-sensitive Kernel Optimization (CsKO) for low-resource on-device incremental learning, boosting accuracy by 38.08% with 75% less computation.

Dingwen Zhang, Yan Li, De Cheng et al.

This study systematically evaluates multiple high-rated LLMs in robotics, revealing significant biases and safety risks, including discrimination and acceptance of unlawful commands.

Andrew Hundt, Rumaisa Azeem, Masoumeh Mansouri et al.

RVT-2 employs multi-stage virtual view reasoning and system-level optimizations to achieve 82% success in high-precision multi-task manipulation with only 10 demonstrations, training 6× faster than RVT.

Ankit Goyal, Valts Blukis, Jie Xu et al.

Real2Code employs image segmentation and fine-tuned LLM code generation to reconstruct complex articulated objects with up to 10 joints, outperforming state-of-the-art methods.

Zhao Mandi, Yijia Weng, Dominik Bauer et al.

GUIOdyssey dataset and OdysseyAgent model enhance cross-app navigation performance.

Quanfeng Lu, Wenqi Shao, Zitao Liu et al.

OLMES standardizes large language model evaluation, covering prompt formatting, example selection, probability normalization, ensuring reproducibility and fairness.

Yuling Gu, Oyvind Tafjord, Bailey Kuehl et al.

cPAPERS dataset integrates OpenReview comments with arXiv documents for multimodal Q&A in scientific papers.

Anirudh Sundar, Jin Xu, William Gay et al.

Proposes a direct synthesis method for ω-regular expressions from transition-based NBAs, yielding more compact expressions, especially effective for LTL formulas.

Charles Pert, Dalal Alrajeh, Alessandra Russo