YOLO26 achieves end-to-end NMS-free detection, removing DFL, with ProgLoss, STAL, and MuSGD, boosting speed and accuracy for edge deployment.

Ranjan Sapkota, Rahul Harsha Cheppally, Ajay Sharda et al.

SPLICE benchmark reveals significant gaps in VLM visual reasoning compared to human performance.

Mohamad Ballout, Okajevo Wilfred, Seyedalireza Yaghoubi et al.

Introduced NeMo task for long video understanding; automated data pipeline; built NeMoBench with 31,378 QA pairs from 13,486 videos.

Zi-Yuan Hu, Shuo Liang, Duo Zheng et al.

Latent Visual Reasoning (LVR) enables end-to-end reasoning in visual embedding space, boosting perception-intensive visual question answering by 5% (71.67% vs 66.67%).

Bangzheng Li, Ximeng Sun, Jiang Liu et al.

MoReact uses diffusion models to generate reactive motions based on textual descriptions, enhancing interaction realism.

Xiyan Xu, Sirui Xu, Yu-Xiong Wang et al.

Introduces ReWatch dataset and multi-stage synthesis with Multi-Agent ReAct, boosting large vision-language models' video reasoning accuracy by over 3%.

Congzhi Zhang, Zhibin Wang, Yinchao Ma et al.

LongLive uses a causal autoregressive model with KV recaching, achieving 20.7FPS for real-time long video generation up to 240 seconds.

Shuai Yang, Wei Huang, Ruihang Chu et al.

JanusVLN employs dual implicit neural memory to separate spatial geometry and semantic info, boosting vision-language navigation.

Shuang Zeng, Dekang Qi, Xinyuan Chang et al.

Proposes a multi-level MOWI framework analyzing hallucinations, revealing data biases and model mechanisms affecting output fidelity.

Zhengyi Ho, Siyuan Liang, Dacheng Tao

Introduces Visual Test-Time Scaling (VTTS) with iterative perception, boosting multimodal reasoning by over 5% on benchmarks.

Ziang Yan, Xinhao Li, Yinan He et al.

Seedream 4.0 uses an efficient diffusion transformer and VAE to enhance high-res multimodal image generation.

Team Seedream, :, Yunpeng Chen et al.

UniPixel unifies pixel-level perception and reasoning, achieving top performance on 10 benchmarks with 3B/7B models.

Ye Liu, Zongyang Ma, Junfu Pu et al.

TempSamp-R1 enhances video temporal grounding via reinforcement fine-tuning, achieving significant accuracy improvements.

Yunheng Li, Jing Cheng, Shaoyong Jia et al.

Do Modern Video-LLMs Need to Listen? By integrating audio encoders, LLaVA-OneVision enhances cross-modal reasoning.

Geewook Kim, Minjoon Seo

SD-VLM enhances VLM's spatial understanding with MSMU dataset and depth positional encoding, outperforming GPT-4o and Intern-VL3-78B.

Pingyi Chen, Yujing Lou, Shen Cao et al.

SPFSplatV2 enables efficient pose-free 3D Gaussian splatting from sparse views, achieving state-of-the-art performance.

Ranran Huang, Krystian Mikolajczyk

Proposed a self-supervised depth estimation method using mixture distribution to enhance boundary sharpness, achieving a 35% improvement.

Aurélien Cecille, Stefan Duffner, Franck Davoine et al.

ScaleCUA leverages cross-platform data to train large-scale GUI agents, achieving 94.4% accuracy on benchmarks.

Zhaoyang Liu, Jingjing Xie, Zichen Ding et al.

GenExam is a multidisciplinary text-to-image benchmark with 1000 samples across 10 subjects, integrating detailed prompts, ground-truth images, and fine-grained scoring.

Zhaokai Wang, Penghao Yin, Xiangyu Zhao et al.

UM-Depth enhances monocular depth estimation with motion and uncertainty, excelling on KITTI datasets.

Tae-Wook Um, Ki-Hyeon Kim, Hyun-Duck Choi et al.