This study conducts a hierarchical attention analysis in VLMs, revealing how middle layers facilitate cross-modal information flow and global feature storage.

Omri Kaduri, Shai Bagon, Tali Dekel

ConsisID introduces a tuning-free, frequency-based control scheme for identity-preserving text-to-video generation using DiT, achieving high fidelity without fine-tuning.

Shenghai Yuan, Jinfa Huang, Xianyi He et al.

CLOVER leverages cross-layer SVD-based orthogonal vectors for pruning and fine-tuning, achieving high compression with minimal performance loss.

Fanxu Meng, Pingzhi Tang, Fan jiang et al.

Proposes 2D Matryoshka training, combining multi-layer and multi-dimension embeddings, significantly improving STS and retrieval performance.

Shuai Wang, Shengyao Zhuang, Bevan Koopman et al.

PersonalVideo achieves high ID-fidelity video generation using reward mechanisms, avoiding dynamic and semantic degradation.

Hengjia Li, Haonan Qiu, Shiwei Zhang et al.

A CBF safety-filter analysis exposes airplane blocking and uses intention-aware adaptive priority to resolve it without communication.

Shuhao Qi, Zengjie Zhang, Zhiyong Sun et al.

PreF3R is a pose-free, feed-forward 3D Gaussian reconstruction method achieving 20FPS for real-time novel-view synthesis.

Zequn Chen, Jiezhi Yang, Heng Yang

This paper proposes a precise scaling law for video diffusion Transformers, integrating optimal hyperparameter prediction to enhance performance and reduce inference costs by 40.1%.

Yuanyang Yin, Yaqi Zhao, Mingwu Zheng et al.

LLM-as-a-judge uses large language models for scoring and selection, enhancing evaluation accuracy.

Dawei Li, Bohan Jiang, Liangjie Huang et al.

Proposes CATP-LLM framework combining TPL and CAORL for cost-aware tool planning, improving plan quality by up to 93.9%.

Duo Wu, Jinghe Wang, Yuan Meng et al.

EPS uses DCT-based spatial-temporal features for efficient patch sampling, reducing training data by up to 91.69% with 82.1× speedup.

Yiying Wei, Hadi Amirpour, Jong Hwan Ko et al.

Zoom Eye enhances multimodal LLMs' visual reasoning with tree search, boosting InternVL2.5-8B by 15.71% on HR-Bench.

Haozhan Shen, Kangjia Zhao, Tiancheng Zhao et al.

FoAR combines high-frequency force/torque sensing with vision, using future contact prediction for dynamic multimodal fusion, greatly improving contact-rich manipulation accuracy.

Zihao He, Hongjie Fang, Jingjing Chen et al.

Introduces orthogonal subspace decomposition via SVD to enhance generalization in AI-generated image detection, outperforming SOTA with >94% AUC.

Zhiyuan Yan, Jiangming Wang, Peng Jin et al.

Defines LLM-as-a-Judge with focus on reliability, bias mitigation, and multi-scenario adaptation, proposing a standardized evaluation framework.

Jiawei Gu, Xuhui Jiang, Zhichao Shi et al.

Using k-SAE to interpret diffusion model features reveals hierarchical semantic information, with intermediate layers capturing fine details.

Dahye Kim, Xavier Thomas, Deepti Ghadiyaram

Proposes truncated diffusion with multi-mode anchors, reducing steps from 20 to 2, achieving 88.1 PDMS at 45FPS for autonomous driving.

Bencheng Liao, Shaoyu Chen, Haoran Yin et al.

XGrammar accelerates CFG-based structured decoding with adaptive caching and persistent stacks, reaching up to 100x lower latency.

Yixin Dong, Charlie F. Ruan, Yaxing Cai et al.

OminiControl leverages parameter reuse of DiT's VAE and Transformer, adding only 0.1% parameters, to achieve versatile multi-task image control surpassing specialized methods.

Zhenxiong Tan, Songhua Liu, Xingyi Yang et al.

FEPS combines active inference with interpretable graph models, avoiding deep neural networks, to enable flexible, transparent decision-making in partially observable environments.

Joséphine Pazem, Marius Krumm, Alexander Q. Vining et al.