Long-Context Transformer Models (LCTMs) are vital for real-world applications but suffer high computational costs due to attention’s quadratic complexity. Block-sparse attention mitigates this by focusing computation on critical regions, yet existing methods struggle with balancing accuracy and efficiency due to costly block importance measurements. In this paper, we introduce XAttention, a plug-and-play framework that dramatically accelerates long-context inference in Transformers models using sparse attention. XAttention’s key innovation is the insight that the sum of antidiagonal values (i.e., from the lower-left to upper-right) in the attention matrix provides a powerful proxy for block importance. This allows for precise identification and pruning of non-essential blocks, resulting in high sparsity and dramatically accelerated inference. Across comprehensive evaluations on demanding long-context benchmarks—including RULER and LongBench for language, VideoMME for video understanding, and VBench for video generation—XAttention achieves accuracy comparable to full attention while delivering substantial computational gains. We demonstrate up to 13.5x acceleration in attention computation. These results underscore XAttention’s ability to unlock the practical potential of block sparse attention, paving the way for scalable and efficient deployment of LCTMs in real-world applications.

Children with Autism Spectrum Disorder (ASD) often exhibit atypical facial expressions. However, the specific objective facial features that underlie this subjective perception remain unclear. In this paper, we introduce a novel dataset, Hugging Rain Man (HRM), which includes facial action units (AUs) manually annotated by FACS experts for both children with ASD and typically developing (TD) children. The dataset comprises a rich collection of posed and spontaneous facial expressions, totaling approximately 130,000 frames, along with 22 AUs, 10 Action Descriptors (ADs), and atypicality ratings. A statistical analysis of static images from the HRM reveals significant differences between the ASD and TD groups across multiple AUs and ADs when displaying the same emotional expressions, confirming that participants with ASD tend to demonstrate more irregular and diverse expression patterns. Subsequently, a temporal regression method was employed to analyze atypicality of dynamic sequences, thereby bridging the gap between subjective perception and objective facial characteristics. Furthermore, baseline results for AU detection are provided for future research reference. This work not only contributes to our understanding of the unique facial expression characteristics associated with ASD but also provides potential tools for ASD early screening. Portions of the dataset and pretrained models are accessible at:https://github.com/Jonas-DL/Hugging-Rain-Man.

Empathy ability is one of the most important social communication skills in early childhood development. To analyze the children's empathy ability, facial expression analysis (FEA) is an effective way due to its ability to understand children's emotional states. Previous works mainly focus on recognizing the facial expression categories yet fail to estimate expression intensity, the latter of which is more important for fine-grained emotion analysis. To this end, this article first proposes to analyze children's empathy ability with both the categories and the intensities of facial expressions. A novel FEA method based on intensity label distribution learning is presented, which aims to recognize expression categories and estimate their intensity levels in an end-to-end framework. First, the intensity label distribution is generated for each frame in the expression sequence using a linear interpolation estimation and a Gaussian function to address the lack of reasonable annotations for expression intensity. Then, the extended intensity label distribution is presented to automatically encode the expression intensity in a multidimensional expression space, which aims to integrate the expression recognition and intensity estimation into a unified framework as well as boost the expression recognition performance by suppressing the variations in appearance caused by intensity and by emphasizing those variations among weak expressions. Finally, a Siamese-like convolutional neural network is presented to learn the expression model from a pair of frames that includes an expressive frame and its corresponding neutral frame using the extended intensity label distribution as the supervised information, thus effectively eliminating the expression-unrelated information's influence on FEA. Numerous experiments validate that the proposed method is promising in analysis of the differences in empathy ability between typically developing children and children with autism spectrum disorder.

Empathy is an important social ability in the early childhood development.One of the significant characteristics of children with autism spectrum disorder (ASD) is their lack of empathy, which makes it difficult for them to feel and understand other people's emotions and to judge other people's behavioral intentions, leading to social disorders.This research designs and implements a facial expression analysis system that could obtain and analyze the real-time facial expressions of children when viewing stimulus materials, and then evaluate the differences of empathy ability between ASD children and typical development (TD) children.The results of this research provide new ideas for the evaluation of ASD children, and also help to develop empathy intervention plans for ASD children.