This paper introduces JudgeAnything, a benchmark leveraging multimodal large language models (MLLMs) as automated judges across 15 diverse tasks, revealing strong performance in understanding but limitations in generation.

Shu Pu, Yaochen Wang, Dongping Chen et al.

OmniLearn dynamically scales worker batches, cutting heterogeneous-cluster training time by 14–85% and improving asynchronous accuracy by up to 6.9%.

Sahil Tyagi, Prateek Sharma

PrediCIR uses a world model to predict missing target features, boosting zero-shot image retrieval by up to 4.45%.

Yuanmin Tang, Jing Yu, Keke Gai et al.

VLU-Net uses a vision-language model for unified multi-degradation image restoration, improving 3.74 dB on the SOTS dehazing dataset.

Haijin Zeng, Xiangming Wang, Yongyong Chen et al.

DCEdit employs dual-level control and precise semantic localization, achieving superior image editing accuracy with no extra training, outperforming existing diffusion-based methods.

Yihan Hu, Jianing Peng, Yiheng Lin et al.

MagicMotion employs dense-to-sparse trajectory guidance within a diffusion transformer framework, outperforming SOTA with a new dataset and benchmark, achieving precise multi-object control.

Quanhao Li, Zhen Xing, Rui Wang et al.

Proposes multi-strategy methods—model pruning, output control, prompt engineering—to reduce reasoning steps, cutting costs while maintaining accuracy.

Yang Sui, Yu-Neng Chuang, Guanchu Wang et al.

This survey systematically analyzes evaluation methods for LLM-based agents, covering core capabilities, application benchmarks, evaluation frameworks, and key dimensions.

Asaf Yehudai, Lilach Eden, Alan Li et al.

ScalingNoise improves video generation quality and consistency by guiding noise selection.

Haolin Yang, Feilong Tang, Ming Hu et al.

CaKE method improves multi-hop reasoning accuracy by 20% on MQuAKE dataset through circuit-aware editing.

Yunzhi Yao, Jizhan Fang, Jia-Chen Gu et al.

HICom uses hybrid-level instruction injection for video token compression, achieving 2.43% performance gain and 78.8% token reduction.

Zhihang Liu, Chen-Wei Xie, Pandeng Li et al.

SpiLiFormer enhances attention allocation in SNNs using lateral inhibition, outperforming SOTA models by 0.46% on ImageNet-1K.

Zeqi Zheng, Yanchen Huang, Yingchao Yu et al.

Proposes Gaussian-HOI optimizer for 3D human-object interaction reconstruction; introduces Open3DHOI dataset with 133 object categories and 120 interactions.

Boran Wen, Dingbang Huang, Zichen Zhang et al.

APEX-MR uses ILP and extended TPG for asynchronous multi-robot cooperative assembly, boosting efficiency by 48% over sequential planning.

Philip Huang, Ruixuan Liu, Shobhit Aggarwal et al.

Introduces ContextualJudgeBench benchmark, evaluating LLM judges' performance in contextual settings, with top accuracy only reaching 55%.

Austin Xu, Srijan Bansal, Yifei Ming et al.

AlignX leverages 1.3 million preference samples to enhance personalized LLM alignment, achieving 17.06% accuracy gain.

Jia-Nan Li, Jian Guan, Songhao Wu et al.

MotionStreamer integrates continuous causal latent space with autoregressive diffusion, enabling real-time streaming human motion synthesis.

Lixing Xiao, Shunlin Lu, Huaijin Pi et al.

EfficientLLaVA employs structural risk minimization for auto-pruning large vision-language models, achieving 83.05% accuracy with only 64 samples and 1.8× speedup.

Yinan Liang, Ziwei Wang, Xiuwei Xu et al.

Proposes TRACE, a semantic-aware transformation method, achieving 30% AUROC improvement for test-time backdoor detection in object detection.

Hangtao Zhang, Yichen Wang, Shihui Yan et al.

POSTA combines diffusion models and multimodal large language models to generate customized artistic posters, covering background, layout, and style control.

Haoyu Chen, Xiaojie Xu, Wenbo Li et al.